TALLYGRAPHAI AGENT SECURITY

Understand the risk behind your AI agents.

Investigate secret exposure and risky sessions in Claude Code and Codex. Analyze transcripts on your machine and use spend to put signals in context.

TERMINAL

# local audit · Claude Code and Codex

$ npx --yes tallygraph trifecta

Reads existing sessions. Does not upload transcripts or change your agent.

From a risky session to an informed decision

Start with a local audit. Review what was observed and what was outside coverage before drawing conclusions.

Audit without connecting an account

Run exposure, trifecta or scan on your machine. Transcripts and secret values stay local.

Review signals and coverage

Tell an observed signal apart from a detector with no data. Partial or unavailable coverage does not mean a session is safe.

Add spend context

Session reports accept --spend to relate usage to risk. Events without a price remain visible as unpriced.

Signals you can investigate

Local audits, session signals and key monitoring have different coverage. A signal is a reason to review, not proof of an attack.

Secret exposure

Find key and token patterns sent to a provider in Claude Code and Codex transcripts. Analysis runs on your machine.

The risky combination

Trifecta identifies sessions that combined private data, untrusted content and a way to send data out. It shows which rules fired.

Suspected injection

Session reports flag instructions in untrusted content followed by a behavior change. These are heuristics to investigate, not confirmed attacks.

MCP tool changes

Compare tool definitions across imported observations. Without a supported observation, coverage is marked unavailable.

Unusual key usage

With Anthropic Admin configured, key watch checks usage every 5 minutes. Spend thresholds and 14 days of history apply, except for decoy keys. It cannot detect every leak.

Secrets before the commit

npx tallygraph scan --staged checks changes staged for commit for secret patterns. The scan runs locally.

Risk and spend, with coverage in view

Cost helps prioritize the review

Relate spend to risky sessions to guide investigation. That spend is not proof of attacker-caused usage.

Sharing requires consent

Team reports are under validation. Each developer decides whether to share metadata with owners; transcripts and secret values stay local.

Observing is not blocking

Session policy runs in shadow mode. Today it reports what it would have decided; no agent has verified enforcement.

SESSION RISKExample
AgentSession spendSignalCoverage
Claude Code$3.40The risky combinationPartial
Codex$1.20Secret exposureComplete
Claude Code$0.80MCP tool changesUnavailable

Start with the sessions on your machine.

Explore the local audit commands. No account needed to check exposure, the trifecta or secrets before a commit.

Already have an account? Sign in here