Talk this topic through with an AI

You get an answer from the documentation and the page it comes from.

leak0 · Reference

leak0 reference

Details on Agentic DLP and Shadow AI for special cases and automation.

~4 min read · leak0 · Reference

Commands

CommandEffect
poison0 leak0 statusShows mode, source (local, profile, default), the number of your own secret files and exclusions, and the protection per agent.
sudo poison0 leak0 enforce, observe, offSets the mode locally.
sudo poison0 leak0 clearRemoves the local setting; then the profile or off applies.
sudo poison0 leak0 file add <glob> --keys k1,k2Adds your own secret file, --keys optional. Also file remove and file list.
sudo poison0 leak0 exclude add <glob>Adds an exclusion. Also exclude remove and exclude list.
poison0 leak0 allow <file>Approves a secret file for commits (10 minutes) and one push.

On Windows, without sudo in an administrator terminal. On cloud devices the changing commands other than clear refuse to run; the console applies there.

Profile fields

FieldValues
leak0.modeenforce, observe or off. New profiles: On, existing ones: Off.
leak0.secretFilesUp to 200 entries { path, keys }; path is a glob, ~ and %USERPROFILE% are allowed.
leak0.exclusionsUp to 200 globs that do not count as a secret source.
leak0.userMayApprovetrue (default) or false: whether users may approve secret files in Git.

Edge cases

SituationBehavior
Token sent to its own sign-in endpoint (Authorization header, user:token@host)Question or note, no block.
Copy to your own servers, S3-compatible storage, your own GitLab or VaultNo block; a question for a credential file.
.env, ~/.zshrc or PEM files in the repositoryQuestion, placeholder or note, no block.
Secret file with real values in a test folderQuestion instead of block.
Checksummed token in a docs folderQuestion instead of block. leak0 checks secret files there like anywhere else.
Git check does not finish in the time availableQuestion, otherwise a block with a request to retry. Never unchecked.

Test folders: test, tests, testdata, fixtures, fixture, __tests__, __fixtures__, __mocks__, plus *_test.*, *.spec.*, *.test.* and templates ending in .example, .sample, .template. Docs folders: docs, doc, examples, example, samples, sample, spec. The path from the repository root counts.

Approval limits

SituationBehavior
New attempt after a refusal or no answerNo new dialog for the same file for 60 seconds, block.
More than three dialogs in one minuteBlock without a dialog.
No desktop, nobody signed in or PoisonZero service not reachableBlock; the message names poison0 leak0 allow.
Failed attempts with allow3 seconds between two attempts, locked for 15 minutes after 5 failures.
File changed after the approvalThe approval no longer applies.

The sign-in of the account the agent runs under counts; no admin rights are needed. The PoisonZero service must be running.

Hooks

SituationBehavior
Installation or turning it on in a cloud profileAdds the hooks to every supported agent found.
leak0 or exec0 offThe hooks stay; only uninstalling removes them.
Hook does not run, Claude CodeThe tool call is blocked when leak0 or exec0 runs in mode enforce.
Hook does not run, other agentsThe agent continues the call.

Audit

ActionMeaning
leak0_blockedCall blocked with proof. Incident.
leak0_askedQuestion asked, for Codex and Gemini CLI with the result of the PoisonZero dialog.
leak0_redactedValue in an output replaced with a placeholder. Incident.
leak0_warnedNote instead of intervention because the agent does not allow it.
leak0_observedHit in mode Observe. In the console only with Also in the console (cloud audit).
leak0_grantedFile approved with poison0 leak0 allow. Incident.
leak0_grant_failedApproval failed, sign-in wrong or cancelled.
leak0_grant_lockedApproval locked after too many failures. Incident.

Fields: agent, tool, rule class, mode and short hash. Never: secret value, path, command or prompt. Short hashes of detected values stay on the device per session and are deleted at the end of the session or after 24 hours.

Not covered

CaseInstead
Memory poisoningMemory protection of the poison0 agent.
Harm without a leak: deleting a credential file, loosening permissionsexec0 checks tool calls for harm.
Prompt injection in content the agent readsinject0 checks fetched content.
Hidden paths through wildcards, variables or $(...)No block; placeholders in the output where the agent allows it.
Re-encoded or split output, such as hexThe whole output is withheld when a credential file is the source.
Scripts, Git hooks or CI that start laterBlock only when the same pipeline sends a credential file to a foreign target.
Passwords without a known formatPlaceholder by field name in known credential files, never a block.
Revoked or test tokens with a valid checksumCannot be told apart offline; a question in test and docs folders.
Paths without a hook: Codex web search, ! commands in Claude Code, cloud runs such as Codex CloudNot covered, no hook sees them.
Plain text in agent transcripts and telemetryNot covered: leak0 changes what the model and foreign hosts see.

Shadow AI: limits and retention

ItemValue
Notice on the deviceText up to 280 characters, link with https:// up to 512 characters, label up to 40 characters
RulesUp to 200 rules, 50 own AI services and 64 agents, up to 20 Google domains
Shadow AI weeksThe last 13 completed weeks
Browser findingsDeleted 30 days after the last report or with the device
Shadow AI turned off for the accountData deleted after 30 days, settings and rules stay

Agents that are not supported

AgentReason
VS Code Agent, Goose, Augment (Auggie), Junie CLI, Factory Droid, Amp, Qwen Code, Kilo Code, Crush, OpenHands, Open Interpreter, gptme, Kimi Code CLI, Mistral Vibe, Pi, Grok CLI, Tabnine CLIHooks exist, not supported in this version.
Cursor, Cline, Cortex Code and Kiro on WindowsNo usable hooks on Windows in this version.
Xcode coding assistantOwn configuration under ~/Library/Developer/Xcode/CodingAssistant/, not protected in this version.
Aider, Zed, Warp, Roo Code, Continue, Trae, Void, Plandex, Qodo, GitHub Copilot in IDEs, JetBrains AI Assistant, Gemini Code Assist, Android Studio Gemini, Databricks Assistant, Snowflake VS Code Extension, Sourcegraph Cody, Letta Code, Tabby, Raycast AI, Google Jules CLI, Pieces, Flowise, Dify, LibreChatNo usable hook known.