agent-safety-review
Skill6 starsChanged 55 days agoMITView on GitHub
Use when a design or implementation involves autonomous agents, unattended/background operation, private-data access combined with external/untrusted input, or outbound actions (sending data, posting, pushing, API calls) — the lethal-trifecta risk
In this repository27 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- agent-team-executionSkill
- agent-team-reviewSkill
- alert-hygieneSkill
- authorial-judgmentSkill
- batch-scriptingSkill
- capture-knowledgeSkill
- deploy-gateSkill
- design-debateSkill
- implementation-drift-checkSkill
- improvement-minerSkill
- incident-analysisSkill
- incident-trend-analyzerSkill
- openspec-shipSkill
- outcome-reviewSkill
- panelSkill
- product-discoverySkill
- project-verificationSkill
- prototype-labSkill
- runtime-validationSkill
- second-opinionSkill
- security-scannerSkill
- skill-scaffoldSkill
- supply-chain-investigationSkill
- synthesizeSkill
- unified-context-stackSkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 6 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈6
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 56
Advisory rule: if the level is execute-reversible or higher AND oversight is weak, emit an advisory proportional to the rung (firmer for …
- Distinctive: line 79
Use synthetic/test data instead of production data
- Distinctive: line 84
Use a quarantine boundary: a read-only agent processes untrusted content, extracts structured data, passes only the structured output to …
- Distinctive: line 90
Use a narrowly scoped HITL: auto-approve low-risk actions, require approval for high-risk ones (sending data externally, deleting …
- Very specific: line 122
When the real threat cannot be exercised in the environment at hand — the sensitive data is absent, the account does not exist, the network …
- Distinctive: line 126
Do not report it as holding.
- Distinctive: line 188
Do NOT claim that improved prompt-injection detection scores solve the problem.
- Distinctive: line 189
Do NOT merge this analysis into security-scanner output.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Agent Safety Review
Architectural risk assessment for designs and implementations that involve autonomous agent behavior. Separate from security-scanner (which runs deterministic static analysis).
When to Use
During DESIGN phase when the prompt involves autonomous agents, unattended operation, private data processing with external input, or outbound actions. Also co-selects during REVIEW phase when autonomy-related triggers match alongside requesting-code-review.
Step 1: Assess the Three Fields
For the proposed design or implementation, evaluate each field:
| Field | Question | Examples |
private_data | Does the agent access information that should not be shared with all parties? | User email, credentials, internal logs, PII, private repos, API keys, session tokens |
untrusted_input | Can an external party inject instructions the agent will process? | Email content, web pages, user-uploaded files, API responses from third parties, webhook payloads |
outbound_action | Can the agent send data or take actions visible outside its sandbox? | Sending emails, posting to Slack, pushing to git, making API calls, writing to shared filesystems, creating PRs |
For each field, state:
- Present — with specific evidence from the design
- Absent — with explanation of why
- Unknown — flag for further investigation
Step 2: Classify Risk
| Fields present | Classification | Action |
| All 3 | Lethal trifecta — High risk | Require mitigation before proceeding |
| 2 of 3 | Elevated risk | Note which leg is missing. Recommend not adding the third without mitigation. |
| 0-1 | Standard risk | No special action required |
Step 2b: Assess Autonomy–Control Coherence
Independent of the trifecta count above — this does not change the trifecta classification; it adds a separate advisory. Data-flow risk and autonomy risk are different axes.
State the agent's intended autonomy level:
| Rung | Meaning |
advise | Proposes; a human acts |
recommend | Proposes a specific action; a human approves each one |
execute-reversible | Acts autonomously; effects bounded, observable, easily undone |
execute-irreversible · unattended | Acts autonomously with hard-to-reverse effects, or runs recurring/unattended with no per-run human checkpoint |
Then state whether proportional oversight is present — strong = per-run approval, HITL checkpoint, manifest+dry-run review, or a bounded/reversible blast radius; weak = none of those.
Advisory rule: if the level is execute-reversible or higher AND
oversight is weak, emit an advisory proportional to the rung (firmer for
execute-irreversible · unattended, softer "consider" for execute-reversible):
autonomy without proportional control is a liability, not power — recommend
designing in per-run approval, narrowing scope, or making actions reversible.
Stay silent when oversight is strong (a reviewed manifest, a required
approval, or reversible/bounded effects already supply the control). This is
guidance, not a veto.
Autonomy levels
The DESIGN-phase hint AUTONOMY CHECK points here. The rungs, the oversight
test and the advisory rule are Step 2b above; at DESIGN time apply them to the
proposed design rather than to a built system, and when the rule fires help the
user design the oversight in before leaving DESIGN. Judge from the actual
design; this is guidance, not a gate.
Step 3: Recommend Mitigation (if lethal trifecta)
The primary mitigation is blast-radius control — cutting at least one leg of the trifecta. Improved detection scores are NOT proof of safety.
Cut private_data:
- Isolate the agent to a sandbox with no access to sensitive data
- Use synthetic/test data instead of production data
- Limit access to only the specific data needed, not broad access
Cut untrusted_input:
- Pre-filter or sanitize external input before the agent processes it
- Use a quarantine boundary: a read-only agent processes untrusted content, extracts structured data, passes only the structured output to the privileged agent
- Restrict input sources to trusted parties only
Cut outbound_action:
- Make the agent read-only — it can analyze but not act
- Require human-in-the-loop approval for all outbound actions
- Use a narrowly scoped HITL: auto-approve low-risk actions, require approval for high-risk ones (sending data externally, deleting resources, creating public artifacts)
Step 3b: Placement check — does the control's region contain the threat?
A control that exists is not a control that applies. For each identified risk, write these down separately, then state whether they overlap:
- Threat region. What the threat is, and which region of the artifact it occupies — which file, which part of that file, which field, which request.
- Control region. Which region the control actually inspects. Not what it is called or intended to cover — what it reads.
- Overlap. State plainly that region 2 contains region 1. If you cannot say that in one sentence, the control does not apply and the risk is unmitigated.
When there is no control to place. This step is about a control you are claiming. A risk you are not claiming any control for needs no placement block — but it must then be reported as unmitigated, in those words. Saying nothing is not the third option: the whole failure this step exists for is a risk that reads as covered because a control was named near it.
Then ask these two questions, in this order:
- What does this check NOT look at?
- Does this check work?
Answer the first in writing before you consider the second. The first is the question that finds a control aimed at the wrong region; the second only ever tests the region the control already inspects, so on its own it certifies a misplaced check.
unvalidated-against
When the real threat cannot be exercised in the environment at hand — the
sensitive data is absent, the account does not exist, the network is
unreachable — record the control as unvalidated-against: <threat>.
Do not report it as holding. Absent evidence is not evidence of absence, and without this clause the gap stops being visible: "the controls stay as designed" silently becomes "the controls hold".
Why this step exists
Measured 2026-09-18. A control was written for a residual risk that had been correctly identified in a trifecta assessment. It hashed one designated element of the artifact against a frozen fixture. The risk lived in a different region of the same file. An artifact carrying a real account identifier in a visible table passed the check and exited 0. Two further bypasses followed: a second data block under another id, and a remote asset reference that fired during capture (proven against a local server — 8 requests before the fix, 0 after).
Nine tasks of red-first testing, mutation verification and adversarial probing ran past it, because every cell varied the contents of the inspected element and nothing probed outside it. The rigor was real and entirely inside the wrong boundary.
Third property, and the one that makes this durable rather than a one-off: the
sensitive data happened to be absent from the machine, so the placement was
unfalsifiable — which is what the unvalidated-against clause above exists
to make visible.
Step 4: Produce Risk Assessment
Output a structured assessment:
## Agent Safety Assessment
**Design:** <what is being evaluated>
**Date:** YYYY-MM-DD
### Risk Fields
| Field | Status | Evidence |
|-------|--------|----------|
| private_data | Present/Absent/Unknown | <specific evidence> |
| untrusted_input | Present/Absent/Unknown | <specific evidence> |
| outbound_action | Present/Absent/Unknown | <specific evidence> |
### Classification
**Risk level:** Lethal trifecta / Elevated / Standard
**Autonomy:** <advise | recommend | execute-reversible | execute-irreversible · unattended> · **Oversight:** <strong | weak>
**Autonomy advisory (if flagged):** <proportional recommendation, or "none">
### Placement (per risk with a claimed control)
- Threat region: <where the threat lives>
- Control region: <what the control inspects>
- Overlap: <one sentence, or "NONE — control does not apply">
- Not looked at: <what this check does not read>
- `unvalidated-against`: <threat that could not be exercised here, or "none">
### Mitigation (if required)
**Recommended approach:** <which leg to cut and how>
**Trade-off:** <what capability is reduced by the mitigation>
**Residual risk:** <what remains after mitigation>Constraints
- This is an architectural review, not a pass/fail gate. The user decides whether to accept the risk.
- Do NOT claim that improved prompt-injection detection scores solve the problem. 97% detection is a failing grade when the 3% leaks private data.
- Do NOT merge this analysis into security-scanner output. Keep architectural risk separate from deterministic code scanning.
- The skill produces an assessment, not a veto. The goal is informed decision-making.
- When the design is an AI/LLM or agent feature, the safety eval cases (injection, escalation, refusal, safety-routing-suppression) MUST be authored and failing (red) before the behavior is implemented — compose with
test-driven-development. Detection added after the behavior exists is not a substitute: a feature that has never failed its safety cases has never been shown to pass them.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

