Free tool · no signup · about 3 minutes
AI Evidence Readiness Checker (Claude)
Whether you can investigate misuse of Claude is decided before the incident, by your plan and by what you switched on. Several sources start recording only from the day they are enabled, and several delete on a rolling clock.
Describe your setup on the left. The checker marks each evidence source available, partial or missing, draws your real retention windows, lists the questions you could not answer today, and orders the fixes. Nothing you enter leaves your browser.
Readiness out of 100
5 missing, 3 partial, 1 available. 2 of 14 investigation questions cannot be answered today.
Shortest window: Claude Code local files, 30 days.
Your three rings
Ring 1
Held by Anthropic
Activity Feed (Compliance API)○ Missing
Not enabled. The feed has no backfill: turning it on during an incident records only from that moment.
claude.ai chat, file and project content○ Missing
The Compliance API is off, so there is no supported route to pull conversations. The CSV export does not contain chat or project content.
Claude Code and Cowork session transcripts○ Missing
The Compliance API is off, so Anthropic is not capturing Claude Code or Cowork session transcripts for you.
Enterprise CSV audit export◐ Partial
Never exported. On demand it looks back 180 days, so activity older than that is already out of reach. With the Compliance API off, this is your only Anthropic-held audit trail.
Ring 2
Telemetry you collect
Claude Code OpenTelemetry○ Missing
Claude Code telemetry is not exported. Host, repository, tool and permission-decision context will not exist for past sessions.
Inference hook security-server records○ Missing
No inference hooks or security server. A supplementary source, so not the first gap to close.
Ring 3
Your endpoints and apps
Claude Code endpoint files◐ Partial
Default cleanupPeriodDays of 30: transcripts and several other paths are removed after a month. Acquire early.
EDR and endpoint security records◐ Partial
EDR telemetry kept 30 days.
Outside the rings
Corroborate outside Claude
Identity provider sign-in logs● Available
IdP sign-in logs kept 90 days.
Your retention timeline
Log scale, one hour to six years. Solid bars are fixed windows; outlined bars are windows you set.
Questions you could not answer today
1. Account compromise · Did someone else get in, and what did they do next?
- ◐Which sign-ins (SSO, social or magic link) did the account make, and from where? Only in part, from CSV audit export, IdP sign-in logs. Full answer needs Activity Feed in good shape.
- ○After the suspicious sign-in, which chats, files and projects did the same actor open or create? Not today. Needs Chat & project content.
- ◐Was the account used to run Claude Code or Cowork sessions, and on which host? Only in part, from Claude Code local files, IdP sign-in logs. Full answer needs Session transcripts, Claude Code OTel in good shape.
2. Persistence & privilege escalation · Did they leave themselves a way back in?
- ◐Were invitations sent, roles assigned or SCIM groups changed? Only in part, from CSV audit export, IdP sign-in logs. Full answer needs Activity Feed in good shape.
- ◐Was an organisation MCP server created or changed, or a desktop extension uploaded? Only in part, from CSV audit export. Full answer needs Activity Feed in good shape.
- ◐Did hooks, plugins, settings or .mcp.json on a developer machine change? Only in part, from Claude Code local files, EDR records. Full answer needs Claude Code local files in good shape.
3. Defense evasion · Did they turn down what you can see or enforce?
- ◐Were IP restrictions, tool policies, the connector domain guard or Cowork settings changed during the window? Only in part, from CSV audit export. Full answer needs Activity Feed in good shape.
- ◐Did anyone switch to bypassPermissions, disable the sandbox or change the permission mode? Only in part, from Claude Code local files. Full answer needs Claude Code OTel, Session transcripts in good shape.
- ◐Was telemetry turned off or pointed somewhere else? Only in part, from Claude Code local files, EDR records. Full answer needs Activity Feed, Claude Code local files in good shape.
4. Data exposure · What reached Claude or a connected system, and where could it go?
- ◐What content did users or the actor put into claude.ai, and was it exported or shared? Only in part, from CSV audit export. Full answer needs Chat & project content in good shape.
- ◐Which files did Claude Code or Cowork read, and what text left the machine? Only in part, from Claude Code local files, EDR records. Full answer needs Session transcripts in good shape.
5. Agent abuse · What did the agent try, who started it, and what actually ran?
- ◐Who started the agent run, and what was it asked to do? Only in part, from Claude Code local files. Full answer needs Session transcripts in good shape.
- ◐Which tool calls were approved, by what source, and what actually executed? Only in part, from Claude Code local files, EDR records. Full answer needs Claude Code OTel in good shape.
- ○Did inference hooks deny requests, fail open or trip a circuit breaker? Not today. Needs Inference hook records.
Fix list, in order
P11. Enable the Compliance API and Activity Feed today
It keeps six years of activity, but only from the day you switch it on. Every day it stays off is a day you can never investigate.
P12. Export Claude Code OpenTelemetry
Gives host, repository, tool and permission-decision context, with tool_use_id and request_id to join to transcripts.
P23. Raise cleanupPeriodDays in managed settings
At 30 days local transcripts vanish before many investigations start. Raise it through managed settings, and remember history.jsonl and memory follow other rules.
P24. Practise the CSV audit export and archive it
It reaches back only 180 days and the download link dies after 24 hours. A scheduled export to storage you own keeps older periods.
P25. Confirm chat and project retention settings
Know how long content is kept and who can delete it, so you can promise an investigator a window.
P36. Consider inference hooks
A security server sees prompts before Claude does and records denials. Supplementary, and it never records the final response.
How Claude evidence is laid out
Picture the incident at the centre of three rings. The outer ring is what Anthropic keeps on your behalf. The Compliance API is the main door: its Activity Feed tells you who signed in, created credentials, changed roles or settings and touched which objects; separate endpoints return retained claude.ai content and Claude Code and Cowork session transcripts. The Enterprise CSV audit export is a narrower fallback, and Claude Platform features such as Files, batches and Managed Agents keep their own data.
The middle ring exists only if you built it. Cowork and Claude Code can export OpenTelemetry to your collector, and Enterprise inference hooks let a security server see prompts first. If nobody pointed telemetry anywhere, there is nothing to recover later.
The inner ring is your own estate: the ~/.claude directory and project files on developer machines, EDR, and the request and response logs of applications you built on the API. The API ring matters because ordinary API prompts and responses are not available through the Compliance API.
The windows that catch people out
- Six years, but only from enablement. The Activity Feed never backfills.
- 180 days and 24 hours. The CSV export looks back at most 180 days, its download link expires a day later, it has no chat or project content, and it is unavailable with customer-managed keys.
- 30 days by default. Claude Code’s cleanupPeriodDays removes local transcripts and several other paths; history and memory files follow other rules.
- Whatever you chose. Chat and project content follows your retention settings; ZDR or HIPAA arrangements can exclude sessions from transcripts; OTel, hook and app logs last as long as your platform keeps them.
One more setting worth checking today: collect content with a key scoped to read:compliance_user_data only. A key with delete scope can permanently remove supported content, which is the last thing you want an intruder, or a hurried analyst, holding.
Questions
What evidence exists if Claude is misused in my organisation?
It sits in three places: records Anthropic keeps for you (the Compliance API with its Activity Feed, retained content and session transcripts, the Enterprise CSV audit export and stored Claude Platform data), telemetry you export yourself (Cowork and Claude Code OpenTelemetry, inference hook records), and files and logs on your own endpoints and applications. Which of these you actually have depends on your plan and what you set up before the incident.
Can I turn on the Activity Feed after an incident and still see what happened?
No. The Activity Feed keeps records for six years from the day it is first enabled, but it does not backfill. Anything that happened before you switched it on is not there, which is why it should be enabled long before you need it.
What changes on the Claude Team plan?
Team has no Compliance API and no CSV audit export. An investigation leans on Claude Code and Cowork OpenTelemetry you configured, files on endpoints, identity provider logs and the records of connected systems.
Are Claude API prompts and responses available through the Compliance API?
No. For applications you build on the Claude API, your own request and response logs may be the only record of what was sent and returned. Log them with the user, session or job identifier that triggered each call.
Keep going
- Evidence Expiry Race: race the retention clocks after an incident is found.
- AI Investigation Scoping Lab: five cases, one per starting point, from sign-in to agent abuse.
- AI Agent Observability Readiness Scorecard: score the wider telemetry around your agents.
- Agent Trace Triage Lab: read Claude OTel spans the way an analyst would.
- AI Incident Response Checklist: the steps once you know an AI tool is involved.
- Where Claude evidence lives: the written guide to these sources.
Further reading: Invictus Incident Response, “Frontier Forensics: Anthropic” (v1.0, October 2026, TLP:CLEAR), the basis for this checker’s source map and retention windows; their cheat sheet is on GitHub. Product behaviour changes: confirm each window in Anthropic’s current documentation before relying on it. This is a self-assessment aid, not legal or compliance advice.
More security tools
The tools people use most on PlayCISO.
Catch injected instructions hiding in content your AI agent should only read.
Score your readiness against the first international standard for AI security threats.
Critical CVEs across 21 major vendors, month by month — and what the 2026 surge means.
Build an ordered ruleset, replay labelled packets and catch shadowed and any-to-any rules.
Self-assess all six Functions, from Govern to Recover, and see your weakest one.
Search 172 threat actors across Microsoft, CrowdStrike, Mandiant and community names.