AI Agent Safety Review

AI Agent Safety Review

Apply discipline per ${CLAUDE_PLUGIN_ROOT}/_shared/DISCIPLINE.md (covers $ARGUMENTS handling, evidence, validation, and safety).

Dispatch policy (V8)

Dispatch target: ultraprompt:auditor (focus: ai-safety). See ${CLAUDE_PLUGIN_ROOT}/_shared/DISPATCH-POLICY.md for the full V8 dispatch decision tree, Task call template, and inline-override conditions.

Distinctive judgment

LLM systems have a unique attack surface: prompts and retrieved content are mixed with instructions, and the model treats both as authoritative. Trust boundaries must be explicit (this is data, that is instruction). Tool-calling is privilege escalation: each tool a model can call is an action it can take. Autonomy controls (human-in-loop, confirmation, audit) are part of the design, not afterthoughts.

First signals to inspect

• System prompt content: are instructions clearly separated from user data?
• Retrieval pipeline: is retrieved content marked as untrusted? Is it shown verbatim to the model?
• Tool inventory: what can the model call? With what side effects?
• Memory/context: what persists across turns or sessions? Who can poison it?
• Autonomy: where does the model act vs ask?
• Output channels: where does model output go? (UI, logs, downstream systems, other models)
• Prompt injection vectors: any path where attacker-controlled text reaches the model

Failure modes specific to this lane

• Concatenating user input directly into the system prompt
• Showing retrieved web content to the model without trust boundary marking (classic prompt injection vector)
• Tool that takes a free-text argument and executes it (shell, eval, SQL)
• Memory shared across users (one user can poison another's session)
• Auto-approval of tool calls that should require human confirmation
• Logging the full prompt + completion (often contains secrets, PII)
• Retrieved context that the model trusts as authoritative when it shouldn't

Workflow

1. Inventory: prompts, tools, retrieval sources, memory, output channels.
2. Map trust boundaries. Confirm separation of instruction vs data.
3. Audit each tool: what does it do, what's the side-effect blast radius, what's the confirmation gate.
4. Audit retrieval: is content marked, sandboxed, or sanitized before reaching the model?
5. Audit memory/context for cross-user contamination.
6. Audit autonomy controls: where can the model act, where must it ask?
7. Audit output channels for log/audit safety.
8. Apply hardening: prompt restructuring, tool argument validation, retrieval marking, confirmation gates.

Validation

Run prompt-injection eval cases (test corpus of known injection strings). Test tool argument validation with adversarial inputs. Test memory isolation across simulated users. For prompts: run before/after eval suite if available.

Output contract

Scope | Trust Boundary Map | Tool Audit (per tool: side effect, confirmation gate, hardening) | Retrieval Audit | Memory/Context Audit | Autonomy Audit | Hardenings Applied | Remaining Risks

Subagent delegation

Dispatch auditor with focus=ai-safety. See _shared/playbooks/prompt-injection-patterns.md and _shared/playbooks/prompt-hardening-checklist.md.

V4 aliases

This skill answers to V4 names: ai-agent-safety-review, prompt-hardening. The router resolves them to ai-agent-safety-review and notes the alias in its response.

Source: Sokoliem/ultraprompt — distributed by TomeVault.