Information Security Risk Assessment // Internal

Unguardrailed LLM Laboratory

A lifecycle risk model for operating locally hosted large language models, including alignment-stripped and agentic configurations, in a research laboratory. Controls mapped to ISO/IEC 27001:2022 Annex A and ISO/IEC 27002:2022.
ClassificationInternal / Restricted
FrameworkISO/IEC 27001:2022 + 27002:2022
Adjacent standardsISO/IEC 42001, 23894, NIST AI RMF
Scoring5×5 qualitative, inherent → residual
Scenarios51 across 10 lifecycle phases
Review cadenceQuarterly, or on material change

00Contents

01Premise and scope

"Zero guardrails" is a statement about the model, not about the system. Removing refusal behaviour from a set of weights does not remove the requirement to control what those weights can reach, what reaches them, what leaves the room, and who is accountable. The guardrail does not disappear; it relocates from the model to the perimeter, the tool boundary, the egress path and the paperwork. A laboratory that accepts unconstrained model behaviour must compensate with materially stronger containment, monitoring, and legal cover than one running a filtered commercial endpoint.

In scope

Out of scope

Assumed architecture baseline

Single-tenant laboratory, one or more GPU hosts, hypervisor or container runtime with GPU passthrough, local vector store, web UI and API endpoints, sited in a domestic or small-office premises with a shared internet connection. Operated by one or a small number of security practitioners who also handle client-confidential material on adjacent systems. This adjacency is the single most consequential architectural fact in the model.

02Risk appetite statement

Risk categoryAppetitePosition
Model output content (offensiveness, refusal absence, harmful text)HIGHAccepted by design. Studying unconstrained behaviour is the purpose of the environment. No content filtering is applied at the model layer.
Research quality and reproducibilityMODERATESome experimental failure is expected. Findings that leave the lab must be verified by a human against primary sources.
Containment failure (egress, lateral movement, host compromise)NONEZero tolerance. Any breakout from the lab boundary is a reportable incident.
Impact on third parties (scanning, exploitation, data exposure)NONEZero tolerance. No lab-originated traffic may reach systems the operator does not own or have written authorisation to test.
Client and personal data exposureNONEZero tolerance. Client material does not enter the lab without documented authorisation and a defined destruction point.
Legal and regulatory breachNONEZero tolerance. Certain generated content categories carry strict liability and are prohibited outright regardless of research intent.
Availability of the lab itselfHIGHNon-critical. Loss of the environment is an inconvenience, not an incident.
Interpretation. The appetite profile is deliberately asymmetric. High tolerance for what the model says; no tolerance for what the system does. Every control in this model exists to enforce that asymmetry.

03Scoring method

Standard 5×5 qualitative assessment. Inherent risk assumes the stated scenario with no laboratory-specific controls applied. Residual risk assumes the mapped controls are implemented and operating effectively. Score = Likelihood × Impact.

Likelihood

LDescriptorGuide
1RareNot expected in the lab's lifetime
2UnlikelyPossible, no known precedent locally
3PossiblePlausible within 12 months
4LikelyExpected within 12 months
5Almost certainOccurring now or on every cycle

Impact

IDescriptorGuide
1NegligibleLocal nuisance, self-correcting
2MinorLost research time, rebuild required
3ModerateFlawed output, minor cost, embarrassment
4MajorClient notification, regulatory interest, material loss
5SevereThird-party harm, criminal exposure, business-ending reputational damage

Heat bands

L\I12345
5510152025
448121620
33691215
2246810
112345

1–4 Low · 5–9 Moderate · 10–14 High · 15–19 Severe · 20–25 Critical

04Lifecycle model

Risk is decomposed across ten phases. Phases 1 and 9 are continuous; the remainder are broadly sequential but iterate.

1 · Governance and authorisation 2 · Supply chain and acquisition 3 · Environment and infrastructure 4 · Data ingestion and preparation 5 · Model adaptation 6 · Runtime and interaction 7 · Agentic operation 8 · Monitoring and incident response 9 · Human factors 10 · Egress, publication and disposal

05Phase 1 — Governance and authorisation

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
GOV-01No documented purpose, authorisation or boundary for the lab. Activity drifts into work that was never sanctioned and cannot be defended after the fact.4416A written Laboratory Authorisation and Acceptable Use standard defines permitted research categories, prohibited categories, the physical and logical boundary, and the named accountable owner. Reviewed annually and on scope change. Every experiment references an authorisation ID.5.1, 5.2, 5.8, 5.10, 5.376
GOV-02No inventory of models, adapters, datasets, agents, prompts or tool definitions. The organisation cannot state what it holds or what produced a given output.4312Asset register covers model artefacts (name, source URI, revision hash, licence, format, quantisation), datasets, adapters, agent configurations and tool manifests. Each artefact carries a classification and a retention decision at registration. Register is the authoritative source for decommissioning.5.9, 5.12, 5.13, 8.96
GOV-03Lab is treated as a personal hobby environment, so ISMS scope, client contractual handling terms and supplier obligations are never applied to it.4416Lab is formally brought inside the ISMS scope statement. A scope decision record documents which client contracts and NDAs constrain lab use. Where a contract prohibits AI processing of engagement data, that prohibition is enforced technically, not by intention.5.1, 5.19–5.22, 5.31, 6.2, 6.68
GOV-04Risk appetite undefined. The decision to run alignment-stripped models is never formally accepted by an accountable person, leaving no defensible record of the trade-off.4312A signed risk acceptance records the deliberate choice to remove model-layer guardrails, the compensating controls relied upon, the expiry date of the acceptance, and the conditions that would trigger reassessment.5.1, 5.35, 5.36, 6.44
GOV-05Sole operator with no independent challenge. Design flaws, control gaps and biased conclusions persist unexamined.4312Scheduled independent review of the lab configuration and its risk model by a competent second party at least annually. Peer review gate on any output intended for publication or client consumption.5.35, 5.36, 6.3, 8.296
GOV-06Professional indemnity or cyber insurance cover invalidated because AI research activity was never declared to the insurer.3412Insurance position reviewed and lab activity disclosed in writing at renewal. Any exclusion for AI-assisted deliverables is recorded on the risk register and reflected in client engagement terms.5.31, 5.19, 5.296
GOV-07Change to lab configuration (new model, new tool, new network path) is made ad hoc with no assessment, silently invalidating the risk position.4312Lightweight change record required for any change to the model set, tool set, network policy or data classification handled. Changes affecting egress or tool privilege require re-approval against the authorisation standard.8.32, 8.9, 5.8, 5.376

06Phase 2 — Supply chain and acquisition

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
SUP-01Malicious serialised model weights achieve code execution on load. Python pickle-based formats deserialise arbitrary objects; loading is equivalent to running untrusted code as the inference user.4520Safetensors or GGUF formats mandated; pickle-based formats prohibited unless there is no alternative. Where unavoidable, first load occurs in a disposable, network-isolated container with no host mounts and no credentials. Static scanning of archives before load. Publisher and revision hash recorded and verified against the register.8.7, 8.8, 8.19, 8.25, 8.22, 5.21, 5.238
SUP-02Weights carry a trigger-conditioned backdoor. Behaviour is benign under evaluation and changes on a specific token, date or context. Static analysis cannot detect this.3412Source restricted to publishers with a verifiable identity and track record. Behavioural evaluation battery run before promotion to research use. Containment is treated as the primary control: the model is assumed potentially adversarial and is denied unmediated egress and tool privilege regardless of evaluation results.5.7, 5.21, 5.22, 8.16, 8.22, 8.2912
SUP-03Typosquatted or impersonated publisher on a model hub delivers a substituted artefact.3412Acquisition procedure requires verification of publisher organisation, pinning to an immutable revision identifier rather than a mutable tag, and recording the artefact digest at download. Digest re-verified before each promotion between environments.5.20, 5.21, 8.9, 8.246
SUP-04Compromised or vulnerable inference dependency in the Python, Node, CUDA or container stack provides initial access.4416Dependency manifests pinned and hash-locked. Software bill of materials maintained for each lab image. Vulnerability feed monitored for the inference runtime and GPU stack specifically. Patch service level defined; images rebuilt rather than patched in place.8.8, 8.9, 8.19, 8.32, 5.7, 5.218
SUP-05Model licence or acceptable use policy breached. An alignment-stripped derivative of a restrictively licensed base model is used to produce commercial client deliverables.4312Licence recorded at registration and enforced by classification. Models with non-commercial or research-only terms are tagged and technically separated from any workflow that touches client deliverables. Derivative and redistribution obligations reviewed before any artefact is shared.5.32, 5.31, 5.9, 5.126
SUP-06Memory-safety defect in a model file parser is exploited by a crafted artefact, giving execution in the inference process.3412Inference process runs unprivileged, in a namespace-isolated container, with a restricted syscall profile and no write access outside a scratch volume. Parser versions tracked in the vulnerability programme. Untrusted artefacts loaded only in the disposable evaluation enclave.8.8, 8.22, 8.27, 8.31, 8.26
SUP-07Weights or tooling sourced from an entity or jurisdiction subject to sanctions or export control, creating a compliance breach on acquisition or onward transfer.248Provenance check at acquisition against sanctions and export control obligations. Onward transfer of any derivative artefact requires an explicit export assessment recorded against the asset register entry.5.31, 5.20, 5.14, 5.94
SUP-08Model card, benchmark claims or capability description materially misrepresent the artefact, leading to unsafe deployment assumptions.428Independent capability and refusal-boundary evaluation performed locally before use. Published claims treated as marketing until verified. Evaluation results attached to the register entry.5.21, 8.29, 8.33, 5.354

07Phase 3 — Environment and infrastructure

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
INF-01Inference API or web UI bound to all interfaces with no authentication and reachable from the wider network or the internet. Default configurations of common inference servers listen without credentials and are routinely enumerated at internet scale.5420All inference and UI services bind to loopback or a dedicated management interface only. Access mediated by an authenticating reverse proxy with strong authentication. No port forwarding or UPnP from the perimeter. External attack surface verified from an off-network vantage point after every change, not assumed.8.3, 8.5, 8.9, 8.20, 8.21, 8.228
INF-02Lab shares a flat network with home or business systems holding client and personal data. Compromise of a GPU host pivots directly to material the operator is contractually obliged to protect.4520Dedicated network segment with default-deny in both directions. No east-west path from lab to trusted segments. Separate DNS and no shared directory service, credential store or file share. Management access is one-way, initiated from the trusted side via a jump host, never the reverse.8.20, 8.22, 8.2, 5.15, 8.18
INF-03Container or hypervisor escape from an inference workload with GPU passthrough. Accelerator passthrough widens the guest-to-host attack surface beyond a standard VM.2510Host runs no other function. Guest is unprivileged with a minimal device surface. Hypervisor and GPU driver stack patched on a defined cycle. Host has no credentials for, and no route to, any trusted segment, so an escape yields an isolated host rather than a foothold.8.8, 8.22, 8.27, 8.31, 7.86
INF-04GPU host hijacked for cryptomining or unauthorised third-party model hosting. High-value accelerators are an explicit targeting objective.339Baseline GPU utilisation, power draw and thermal profile established. Alerting on sustained deviation outside scheduled workloads. Endpoint protection and egress monitoring on the host. Scheduled review of running processes and container inventory.8.6, 8.7, 8.16, 8.194
INF-05Remote access left exposed. A convenience tunnel, remote desktop service or SSH endpoint stood up for a single experiment is never torn down.4520Remote access exclusively via the sanctioned VPN with phishing-resistant multi-factor authentication. Ad hoc tunnelling utilities prohibited on lab hosts and blocked at egress. Temporary access provisioned with an expiry and reconciled weekly against an external surface scan.6.7, 8.5, 8.20, 8.21, 8.16, 8.198
INF-06Physical theft of high-value accelerators or storage containing models, logs and research data.248Equipment sited in a lockable area with controlled entry. Full disk encryption on all lab storage with keys not resident on the device. Asset serial numbers recorded. Loss triggers the incident procedure including assessment of any client data resident at the time.7.1–7.4, 7.8, 7.9, 8.24, 5.244
INF-07Thermal, electrical or fire risk from sustained multi-GPU load in a domestic setting. Server-class passively cooled accelerators repurposed into consumer chassis, combined with marginal power supply headroom and shared ring-main circuits, is a physical safety issue, not just an availability one.3515Documented thermal and power design with measured headroom under sustained full load. Active cooling verified for passively cooled cards. Circuit loading assessed against the supply. Interlinked smoke detection covering the equipment location. No unattended long-duration runs without automated thermal shutdown thresholds and remote power isolation.7.5, 7.8, 7.11, 7.12, 7.13, 5.298
INF-08Capacity exhaustion. Model artefacts and checkpoints consume storage until the host destabilises and research data is lost.428Storage and VRAM capacity monitored with thresholds. Artefact retention enforced from the register. Research outputs backed up independently of the working volume.8.6, 8.13, 8.164
INF-09Unpatched supporting infrastructure in the lab estate: hypervisor management interface, storage appliance, router, IPMI or BMC.4416Full asset coverage in the vulnerability management programme, explicitly including management planes. Out-of-band management interfaces isolated to a dedicated segment with no internet route. Default credentials eliminated at build.8.8, 8.9, 8.20, 8.22, 5.178
INF-10Lab data backed up to consumer cloud storage with no encryption, no contract and no processing terms.3412Backup destinations approved and recorded. Client-encrypted before transfer with keys held outside the provider. Where the backup may contain personal data, provider terms assessed against data protection obligations before use.8.13, 8.24, 5.23, 5.34, 5.146

08Phase 4 — Data ingestion and preparation

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
DAT-01Client-confidential material (test evidence, incident artefacts, source-restricted intelligence) is ingested for retrieval or fine-tuning, breaching contractual handling terms and non-disclosure obligations.4520Default position is that no client data enters the lab. Exceptions require written client authorisation naming the environment, the purpose, the retention period and the destruction method. Classification labels enforced at the ingestion boundary. Data loss prevention on transfer paths into the lab segment. Sanitised or synthetic substitutes used wherever the research question permits.5.12, 5.13, 5.14, 5.34, 6.6, 8.3, 8.11, 8.1210
DAT-02Personal data processed with no lawful basis, no impact assessment and no minimisation. Model training and retrieval are high-risk processing under data protection law.4416Data protection impact assessment completed for the lab as a processing environment and refreshed on material change. Lawful basis and minimisation documented per dataset. Special category data prohibited unless a specific assessed exception applies. Records of processing updated to include the lab.5.34, 5.31, 5.33, 8.10, 8.118
DAT-03Cross-contamination between engagements via a shared vector store, shared conversation history or shared adapter, so material from one client surfaces in work for another.3515One retrieval index and one storage namespace per engagement, never shared. Namespace destroyed at engagement close, with destruction evidenced. Access restricted per namespace. Adapters trained on engagement data are treated as carrying that data's classification.8.3, 8.10, 8.31, 5.12, 5.338
DAT-04Live malicious content ingested into a retrieval corpus detonates in the parsing chain rather than in the model. Document parsers, archive extractors and image libraries are the execution surface, and security research corpora are exactly where such content lives.4520All ingestion parsing occurs in a dedicated, network-isolated, ephemeral sandbox distinct from the inference host. Output of parsing is text only, with active content stripped. Archive extraction bounded for depth, ratio and size. Malware analysis material handled under the separate malware handling procedure and never ingested directly.8.7, 8.22, 8.27, 8.31, 8.198
DAT-05Corpus poisoning. Attacker-controlled or opportunistically scraped content plants instructions that alter downstream model behaviour, or shifts a fine-tune's behaviour in a way that is not visible at evaluation time.4416Corpus provenance recorded per source with a trust rating. Untrusted scraped content is segregated into a distinct index and never mixed with trusted sources. Content integrity checked at ingestion. Retrieved content is treated as untrusted data at inference time, never as instruction.5.7, 8.29, 8.33, 8.16, 5.1210
DAT-06Sensitive data persists beyond intent in conversation history, embeddings, cached key-value state, swap files, temporary directories and application logs.4416Retention defined per data store including caches and history. Automated purge on schedule and at engagement close. Swap disabled or encrypted on inference hosts. Deletion verified rather than assumed, with evidence retained against the register entry.8.10, 8.24, 8.15, 5.33, 5.128
DAT-07Copyright or licence-restricted corpora used in a derivative work that is then published or delivered commercially.339Source licence recorded per corpus. Publication gate checks the licence position of every contributing source. Rights-restricted material excluded from any corpus feeding deliverable production.5.32, 5.31, 5.94
DAT-08Embedding inversion or membership inference recovers source text from a vector store that was assumed to be an opaque numerical artefact.248Vector stores classified at the same level as their source corpus, not lower. Encrypted at rest, access-controlled, and never shared or published on the assumption that embeddings are non-sensitive.5.12, 8.3, 8.24, 5.144

09Phase 5 — Model adaptation and alignment modification

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
ADP-01Deliberate alignment removal produces an artefact with meaningful uplift for harmful tasks. The artefact leaks, is shared, or is redistributed, and the lab becomes the origin of a capability now in circulation.3515Alignment-modified artefacts are classified at the highest internal level and marked non-transferable. Stored encrypted on the lab segment only, never on portable media, endpoints or cloud sync. Egress of model files blocked by policy and by data loss prevention. Any proposed external sharing requires documented senior approval and an assessment against the acceptable use standard. Destruction is mandatory at project close unless a retention decision is recorded.5.10, 5.12, 5.13, 5.14, 8.3, 8.10, 8.12, 8.24, 6.610
ADP-02Fine-tuning causes memorisation. The model reproduces training material verbatim, including sensitive content the operator believed was abstracted away by the training process.3412Training data minimised and masked before use. Post-training extraction testing performed against known sensitive strings. Models trained on classified material inherit that classification permanently and are not promoted to lower-trust use.8.11, 8.29, 8.33, 5.126
ADP-03Adapter and checkpoint sprawl. No provenance chain linking an output back to the exact weights, adapter, quantisation, prompt and sampling parameters that produced it. Research is unreproducible and findings are indefensible.4312Every generated artefact records base model digest, adapter digest, quantisation, runtime version, system prompt and sampling parameters. Experiment records retained under the records control. Untracked checkpoints are deleted on discovery.5.9, 5.33, 8.9, 8.32, 5.376
ADP-04Adaptation silently degrades capability or introduces bias. Downstream research conclusions are wrong in a way that is not obvious.339Pre- and post-adaptation evaluation against a fixed held-out benchmark set. Regression threshold defined; artefacts failing the threshold are not promoted. Evaluation set held separately from training data.8.29, 8.33, 8.324
ADP-05Unattended multi-day training run with no supervision of thermal state, integrity or completion.339Long-running jobs require automated monitoring with thermal and failure alerting to a channel the operator actually watches. Checkpointing enabled so a failure is a delay, not a loss. Remote power isolation available.8.6, 8.16, 7.5, 5.294

10Phase 6 — Runtime and interaction

Framing note. In a guardrailed deployment, jailbreak resistance is a control. Here it is absent by design, so it cannot be counted as one. Every scenario in this phase is assessed on the assumption that the model will comply with any instruction it receives, from any source, including instructions embedded in data it was merely asked to read.
IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
RUN-01Direct prompt manipulation steers the model into behaviour outside the sanctioned research scope. There is no refusal boundary to fall back on.5315Accepted at the model layer. Compensating controls are environmental: the model has no privileged tool access by default, no unmediated egress, and full session logging. Scope enforcement is procedural via the acceptable use standard and periodic log review, not technical refusal.5.10, 8.15, 8.16, 8.20, 8.22, 6.312
RUN-02Indirect prompt injection. Instructions embedded in fetched web pages, documents, code comments, filenames or tool output are executed as though the operator had issued them.5525Architectural separation of instruction and data channels. Retrieved and fetched content is delimited, marked as untrusted, and never granted authority. Web access restricted to an allowlist through a filtering proxy, or removed entirely for sessions handling untrusted corpora. Tool invocation triggered by content that entered from an untrusted channel requires human confirmation. Full tool-call logging with alerting on unexpected tool sequences.8.20, 8.22, 8.23, 8.26, 8.27, 8.16, 5.712
RUN-03Insecure output handling. Model output is passed to a shell, an evaluation function, a database query or a browser render without validation, converting a text prediction into code execution or injection.4520Model output is treated as untrusted input to every downstream consumer, with the same validation, encoding and parameterisation applied as to any external data. Automatic execution of generated code is prohibited outside an ephemeral, network-isolated sandbox. Rendered output is contextually encoded. Secure coding standards apply to all lab-built tooling.8.25, 8.26, 8.27, 8.28, 8.298
RUN-04Fabricated technical content reaches a client deliverable or a public publication. Plausible but invented indicators, CVE detail, attribution claims or code carry the operator's professional credibility behind them.4416Mandatory human verification gate: no model-generated technical assertion leaves the lab without independent confirmation against a primary source. Provenance of AI assistance recorded internally and disclosed where client terms require. Attribution claims and indicators subject to a separate confidence-assessment procedure. Peer review before publication.5.35, 5.36, 5.37, 6.3, 8.29, 5.318
RUN-05Generation of material that is unlawful to create or possess. Multimodal capability plus zero refusal behaviour plus automatic caching means an offence can be committed and evidenced on disk without deliberate intent. Several relevant offences are strict liability, and research purpose is not a defence.2510Absolute prohibition, stated explicitly in the acceptable use standard and the risk acceptance, on prompting toward prohibited content categories. Multimodal generation capability disabled unless a specific authorised research need exists. Output caching and history disabled for any session with generative image capability. A written procedure defines immediate actions on inadvertent generation: isolate, do not distribute, do not attempt independent deletion of evidence, obtain legal advice, report. Awareness training covers the strict-liability position directly.5.10, 5.31, 5.24, 5.28, 6.3, 8.10, 8.2310
RUN-06Resource exhaustion through adversarial context length or recursive generation, denying the lab to legitimate use and driving thermal load.326Context length, token budget and concurrency limits enforced at the serving layer. Per-session quotas. Utilisation alerting.8.6, 8.16, 8.213
RUN-07Credentials pasted into prompts. API keys, client credentials and connection strings entered for convenience persist in conversation history, cache and logs indefinitely.4416Secrets are supplied to tooling through a secrets manager and environment injection, never through the conversation channel. Pattern-based detection on prompt input and stored history with alerting. History and log purge procedure on detection, followed by credential rotation as the primary remediation.5.17, 8.12, 8.15, 8.24, 8.10, 6.38
RUN-08Model extraction or inversion against an exposed or over-shared endpoint recovers a proprietary fine-tune or its training characteristics.236Endpoints authenticated and rate-limited. Access to adapted models restricted to named identities. Query volume monitored for extraction-shaped patterns.8.3, 8.5, 8.16, 8.213

11Phase 7 — Agentic operation

Highest-consequence phase. An unguardrailed model with tool access is not a text generator with a plugin; it is an unauthenticated remote code execution primitive that anyone able to influence its input can reach. This phase carries the model's only credible route to third-party harm and criminal liability, and it warrants the strictest controls in the document.
IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
AGT-01Excessive agency. The agent holds shell, filesystem and network tools with no approval gate, so any injected instruction becomes host command execution under the agent's privileges.4520Least-privilege tool design: each tool exposes the narrowest possible operation rather than a general-purpose shell. Agent runs as an unprivileged identity in an ephemeral container with no host mount and no credentials beyond the task. Human approval required for any state-changing, destructive or outbound action. Privileged utility programs removed from the agent image. Session-scoped, time-bounded tokens only.8.2, 8.3, 8.18, 8.22, 8.27, 5.15, 5.1712
AGT-02The agent initiates scanning, enumeration or exploitation against systems the operator has no authorisation to test, from the operator's own IP address. This is a criminal exposure under computer misuse legislation and a breach of connectivity terms, and intent is not a required element of the primary offence.3515Default-deny egress at the lab boundary. No default route for agent workloads. Any external destination requires explicit allowlisting with a recorded justification and an expiry. Offensive tooling absent from the agent image entirely. Where external interaction is genuinely required for research, it targets operator-owned infrastructure or a documented, authorised range only. Egress logging retained and reviewed.8.20, 8.22, 8.23, 8.16, 5.10, 5.3110
AGT-03The agent exfiltrates lab or client data through a channel that was permitted for legitimate reasons: a webhook, a repository push, a paste service, a DNS lookup or an image request.3515Egress allowlist covers destination and protocol, with DNS forced through a controlled resolver and logged. Outbound repository, messaging and paste destinations blocked by default. Data loss prevention on permitted paths. Volumetric and entropy alerting on outbound flows from lab hosts.8.12, 8.20, 8.21, 8.23, 8.16, 5.1410
AGT-04The agent harvests credentials available in its execution context: environment variables, key material, cloud tokens, browser credential stores, configuration files.3515Agent execution context contains no credentials other than the minimum for the specific task, issued short-lived from a secrets manager. No user profile, browser data or key material present in the agent image. Filesystem access restricted to a task-scoped scratch volume. Credential use monitored for anomalous source or pattern.5.17, 8.2, 8.3, 8.24, 8.16, 8.188
AGT-05The agent establishes persistence or modifies its own configuration: scheduled tasks, service units, edits to its own system prompt or tool manifest.3412Agent runtime is immutable and rebuilt from a controlled image each session; no state survives by default. Prompt and tool definitions held in version control outside the agent's write path. File integrity monitoring on configuration and scheduling locations. Software installation on lab hosts restricted.8.9, 8.19, 8.32, 8.16, 8.46
AGT-06Unbounded agent loop consumes compute, power and external API spend without producing useful output.428Hard limits on iteration count, wall-clock duration, token spend and tool invocations per session. Automatic termination at threshold. Spend alerting on any metered external service.8.6, 8.16, 5.304
AGT-07Confused deputy. The agent exercises its own privileges on behalf of a request that originated from an untrusted source, so authority is laundered through the agent.3412Requester identity propagated to the authorisation decision rather than the agent acting under a single service identity. Distinct, lower-privilege agent profiles for sessions that touch untrusted content. Authorisation evaluated at the tool boundary, not the prompt boundary.8.2, 8.3, 8.27, 5.15, 5.168
AGT-08Multi-agent or long-running autonomous configuration behaves outside its expected envelope and no reliable stop mechanism exists.248A tested kill switch exists at three layers: process termination, network isolation of the segment, and physical power isolation. Termination procedure documented and exercised. Autonomous runs are attended or bounded by an automatic expiry.5.29, 5.30, 8.16, 8.27, 5.374

12Phase 8 — Monitoring and incident response

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
MON-01No logging of prompts, responses, tool calls or model load events. An incident cannot be reconstructed, and the operator cannot demonstrate what the lab did or did not do.4416Structured logging of session identity, model and adapter digests, prompts, responses, tool invocations with arguments and results, and artefact load events. Logs shipped off-host in near real time to a store the lab identity cannot write to. Clock synchronisation enforced across lab hosts.8.15, 8.16, 8.17, 5.286
MON-02The logs become the largest single concentration of sensitive material in the environment, and are protected less well than the systems they describe.4416Prompt and response logs classified at the highest classification of any data the lab handles. Encrypted at rest, access restricted to named identities, retention bounded and enforced. Secret and personal data redaction applied at ingestion where feasible.8.15, 8.24, 8.11, 8.10, 5.33, 5.128
MON-03No detection content specific to lab behaviour. Generic endpoint alerting does not distinguish legitimate research from compromise or agent misbehaviour.4312Detection use cases developed for lab-specific signals: unexpected model artefact writes, egress to non-allowlisted destinations, tool invocation outside baseline sequences, GPU utilisation anomalies, new listening services, credential access from agent contexts. Use cases tested against simulated conditions, not assumed.8.16, 5.7, 8.296
MON-04No incident response plan covering lab-specific events, including who is notified and within what deadline when client data or third parties are involved.3412Lab-specific scenarios added to the incident response plan: model artefact compromise, agent egress event, inadvertent unlawful content generation, client data exposure. Notification obligations mapped per scenario, including regulatory and contractual deadlines. Evidence preservation procedure defined. Plan exercised at least annually.5.24–5.28, 5.31, 5.34, 5.296
MON-05Log integrity failure. An attacker, or the agent itself, clears or alters lab logs and removes the evidence of its own activity.339Append-only remote log store. Lab hosts hold write-only credentials to the collector and no deletion capability. Integrity verification on the archive. Alerting on log source silence.8.15, 8.24, 8.22, 5.33, 8.164

13Phase 9 — Human factors

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
HUM-01Automation bias. Fluent, confident output from an unaligned model is trusted because it reads as authoritative, and anthropomorphic framing of the system encourages the operator to attribute reasoning and intent where there is none.4416Awareness training covering the limits of generative output, the distinction between fluency and accuracy, and the specific hazard of anthropomorphic language in technical judgement. Verification gate on all technical assertions. Documented procedure requiring the operator to record the primary source for any model-originated claim used in analysis.6.3, 5.37, 5.35, 5.368
HUM-02Unauthorised physical or logical access by household members, guests, cleaners or contractors to a lab holding client material and unfiltered models.339Lab area secured when unattended. Sessions locked on idle. Named accounts with strong authentication; no shared or unauthenticated local access. Clear desk and clear screen applied to lab equipment. Visitor access to the area recorded.7.1–7.4, 7.7, 8.1, 8.5, 5.154
HUM-03Operator exposure to distressing generated content with no welfare provision, in an environment specifically configured not to refuse.339Welfare provision documented as part of the authorisation standard, including exposure limits, a route to support, and an explicit position that operators may stop any line of research without justification. Prohibited content categories stated so the boundary is known in advance rather than discovered.6.3, 6.8, 5.10, 6.26
HUM-04Key person dependency. The environment is undocumented and only one person can operate, secure or safely shut it down.4312Documented operating procedures covering build, operation, emergency shutdown and decommissioning, held outside the lab and readable by a competent third party. Named deputy with documented access to isolation and power controls. Configuration held in version control.5.2, 5.37, 5.29, 8.13, 8.96
HUM-05Research findings published without ethical review, responsible disclosure consideration or assessment of whether the publication itself provides uplift.3412Publication gate requiring assessment of uplift, disclosure obligations to affected parties, licence and copyright position, and client confidentiality. Peer review before release. Documented decision record retained for each publication.5.35, 5.14, 5.31, 5.32, 6.2, 5.56

14Phase 10 — Egress, publication and disposal

IDRisk scenarioLIInhCountermeasures and control statementsISO 27002:2022Res
EOL-01An alignment-modified artefact, a harmful output sample or a working exploitation capability is shared publicly with no gating, because sharing is the cultural default in the research community.3515Information transfer control applies to model artefacts as information assets. Alignment-modified weights are non-transferable by classification. Publication gate covers uplift assessment, licence position and disclosure obligations. Repository and model hub destinations blocked from lab egress by default.5.14, 5.10, 5.12, 8.12, 8.23, 5.3510
EOL-02Decommissioned accelerators, drives or hosts are sold or returned under warranty with residual model artefacts, logs and client data recoverable.3412Secure disposal procedure covering cryptographic erasure, verified media sanitisation and certificate of destruction. Full disk encryption from build so key destruction is the primary mechanism. Disposal recorded against the asset register entry before the asset leaves the premises.7.10, 7.14, 8.10, 8.24, 5.94
EOL-03Backup, snapshot and image sprawl retains data long past its retention decision, defeating deletion at the primary store.4312Retention applied to backups and snapshots equally, with automated expiry. Deletion requests traced through every copy including offline media. Backup inventory reconciled against the asset register quarterly.8.13, 8.10, 5.33, 5.96
EOL-04The lab is abandoned rather than decommissioned: left powered, reachable and unpatched after the research it was built for has finished.3412Every lab instance carries a defined end date at authorisation. Decommissioning procedure covers data destruction, network removal, credential revocation and register closure. Quarterly reconciliation of running assets against active authorisations, with unauthorised survivors powered down.5.9, 5.37, 8.9, 8.16, 8.84

15Consolidated control statements

Statement of Applicability extract. These are the controls whose implementation is materially altered by the presence of an unguardrailed LLM laboratory, expressed as auditable statements.

RefControlLaboratory-specific implementation statement
A.5.1Policies for information securityA Laboratory Authorisation and Acceptable Use standard is approved and maintained. It states permitted and prohibited research categories, the accepted position on alignment-stripped models, the compensating controls relied upon, and the conditions under which the acceptance lapses.
A.5.9 / A.5.12Asset inventory and classificationModel weights, adapters, prompts, tool manifests, corpora and vector stores are registered as information assets with source, digest, licence and classification. Alignment-modified artefacts and prompt logs are classified at the highest internal level. Vector stores inherit the classification of their source corpus.
A.5.14Information transferModel artefacts are subject to transfer control as information assets. Alignment-modified weights are marked non-transferable. Any external transfer requires documented approval, an uplift assessment and a licence check.
A.5.17Authentication informationNo credential is supplied to a model or agent through the conversation channel. Agent credentials are short-lived, task-scoped and issued from a secrets manager. Detection of a credential in prompt or log content triggers rotation as the primary remediation.
A.5.31 / A.5.32Legal, regulatory and intellectual propertyLegal obligations applicable to the lab are identified and maintained, covering computer misuse, data protection, unlawful content offences, export control and model licensing. Prohibited content categories are stated absolutely, with no research exemption asserted.
A.5.34Privacy and PII protectionA data protection impact assessment covers the lab as a processing environment. Personal data ingestion requires a documented lawful basis, minimisation and a destruction point. Special category data is prohibited absent a specific assessed exception.
A.6.3Awareness and trainingOperators are trained on prompt injection, insecure output handling, automation bias, anthropomorphisation risk, the strict-liability position on certain generated content, and the incident procedure for inadvertent generation.
A.7.5 / A.7.11Physical and environmental threats, utilitiesThermal and electrical loading of the accelerator estate is calculated and measured under sustained full load. Active cooling is verified for server-class passive cards. Interlinked smoke detection covers the equipment location. Automated thermal shutdown thresholds are configured for unattended runs.
A.8.2 / A.8.18Privileged access and utility programsAgents execute as unprivileged identities. Tools expose narrow operations rather than general-purpose shells. Privileged utilities and offensive tooling are absent from agent images. State-changing and outbound actions require human approval.
A.8.10Information deletionRetention is defined and enforced for conversation history, caches, embeddings, temporary artefacts and logs. Deletion is verified rather than assumed, and traced through backups and snapshots.
A.8.15 / A.8.16Logging and monitoringSession identity, model digest, prompts, responses, tool calls and artefact loads are logged and shipped to an append-only store the lab cannot write to. Detection content exists for lab-specific signals and is tested rather than assumed effective.
A.8.20 / A.8.22 / A.8.23Network security, segregation and filteringThe lab occupies a dedicated segment with default-deny in both directions and no default route for agent workloads. External destinations are allowlisted with justification and expiry. DNS is forced through a logged resolver. Management access is initiated only from the trusted side.
A.8.26 / A.8.27Application security and secure architectureInstruction and data channels are architecturally separated. Retrieved content is delimited and carries no authority. Authorisation decisions occur at the tool boundary, not the prompt boundary. Untrusted-content sessions run under a reduced-privilege agent profile.
A.8.28Secure codingModel output is treated as untrusted input by every downstream consumer, with validation, contextual encoding and parameterised queries applied. Automatic execution of generated code occurs only in an ephemeral, network-isolated sandbox.
A.8.29 / A.8.33Security testing and test informationModels are evaluated locally for capability and refusal boundary before promotion. Adaptation is gated on a regression threshold against a held-out set. Sanitised or synthetic data is used in place of live client data wherever the research question permits.
A.8.31Separation of environmentsArtefact evaluation, corpus parsing, inference and agent execution occur in separate isolated contexts. One retrieval namespace per engagement, destroyed at close with evidence retained.
A.5.29 / A.5.30Continuity and readinessA tested kill switch exists at process, network and physical power layers. The termination procedure is documented and exercised. Autonomous runs are attended or bounded by automatic expiry.

16Residual risk view

Distribution shift

BandInherentResidualMovement
Critical (20–25)110−11
Severe (15–19)80−8
High (10–14)219−12
Moderate (5–9)1128+17
Low (1–4)014+14
Total5151

Residual High-band risks requiring active management

IDResidual exposureResWhy it does not reduce further
RUN-02Indirect prompt injection12No general defence exists. Instruction and data share one channel by architecture. Controls constrain blast radius, not occurrence.
AGT-01Excessive agency12Tool access is the research objective. Reducing it further removes the capability being studied. Approval gates depend on human vigilance, which degrades with volume.
RUN-01Direct prompt manipulation12Accepted by design. Refusal behaviour is deliberately absent; the residual is the cost of the research premise.
SUP-02Trigger-conditioned model backdoor12Not detectable by static or behavioural evaluation with current methods. Managed purely by containment.
DAT-01Client data ingestion10Depends on operator discipline at the moment of convenience. Technical prevention is partial where the operator holds legitimate access to both environments.
ADP-01Alignment-stripped artefact leakage10Impact remains severe under any likelihood reduction. Controls reduce probability but cannot reduce consequence.
AGT-02Unauthorised outbound activity10Criminal exposure keeps impact at maximum. Egress controls are strong but a single misconfiguration is sufficient.
AGT-03Agent exfiltration via permitted channels10Any permitted egress path is a potential covert channel. Complete closure would prevent legitimate research.
RUN-05Unlawful content generation10Strict liability means impact cannot be mitigated, only likelihood. Prohibition plus capability removal is the entire control set.
DAT-05Corpus poisoning10Detection of semantic poisoning in large corpora is not reliable. Segregation limits contamination rather than preventing it.
EOL-01Publication of harmful artefacts10Governed by human judgement at the publication gate. Technical egress blocking is bypassable by an authorised operator.

17Irreducible risk and formal acceptance

Three residual risks cannot be engineered away and must be accepted explicitly or the lab should not operate.

These acceptances should be signed, dated and given an expiry no longer than twelve months. An acceptance without an expiry is not a risk decision, it is an omission with a signature on it.

InstrumentRelevance to the labPractical consequence
Computer Misuse Act 1990AGT-02. Unauthorised access to a computer is the offence; an agent acting autonomously does not remove the operator's exposure.Default-deny egress is a legal control, not merely a technical one. Authorisation for any external testing must be in writing and held before the fact.
UK GDPR and Data Protection Act 2018DAT-01, DAT-02, DAT-06, MON-02. Model training and retrieval over personal data is high-risk processing.Impact assessment required. Records of processing must include the lab. Deletion must be demonstrable across caches, embeddings and logs.
Protection of Children Act 1978 and related offencesRUN-05. Generation and possession offences are strict liability; multimodal capability plus automatic caching creates exposure without intent.Disable multimodal generation absent specific need. Disable caching for any session with that capability. Rehearse the response procedure.
Terrorism Act 2000 and 2006RUN-05, EOL-01. Collection and dissemination offences apply to generated as well as collected material.Prohibited categories stated absolutely in the acceptable use standard. Publication gate assesses dissemination risk.
Model licences and acceptable use policiesSUP-05, ADP-01. Many open-weight licences prohibit alignment removal, certain use categories, or commercial derivative use.Licence recorded at registration; commercial workflows technically separated from restricted artefacts.
Export control and sanctionsSUP-07, EOL-01. Weights and certain tooling may be controlled on transfer or by counterparty.Provenance check at acquisition; export assessment before any onward transfer.
Client contracts and NDAsGOV-03, DAT-01, DAT-03. Many engagement terms prohibit processing of client data by AI systems or require prior consent.Contractual position mapped per client. Where prohibited, enforce technically rather than by intention.
Connectivity terms of serviceAGT-02, INF-05. Domestic and business connectivity terms typically prohibit scanning and hosting.Egress allowlist. No inbound exposure of lab services.

19Where ISO 27001 runs out

ISO/IEC 27001:2022 Annex A contains no AI-specific control. Every mapping in this document reuses a general control in an AI context, which works well for containment, access, logging and supply chain, and works poorly for the risks that are genuinely novel. Four areas have no clean Annex A home:

Recommended supplementary alignment. Use ISO/IEC 27001 as the containment and assurance backbone, then layer AI-specific management on top rather than stretching Annex A to cover it.

ISO/IEC 42001 AI management system: impact assessment, oversight, lifecycle governance.
ISO/IEC 23894 AI risk management guidance, maps cleanly onto this register's structure.
NIST AI RMF Govern, Map, Measure, Manage functions for the assurance gap.
OWASP Top 10 for LLM Applications Direct technical mapping for RUN-02, RUN-03, RUN-07, AGT-01.
MITRE ATLAS Adversarial technique taxonomy for detection content development under MON-03.
NCSC secure AI system development guidance Secure design, development, deployment and operation principles.

Assurance and review