What changed, and what was measured.

Latest verified PyPI release: AgentLeak 0.20.1, published 2026-11-16. Release catalog checked 2026-09-28.

The bundled-corpus benchmark is a separate snapshot: 491 scenarios measured on 2026-11-15 with AgentLeak 0.20.0. A newer release does not silently update those measurements. The research paper is versioned independently.

latest0.16.00.17.00.18.00.19.00.19.10.20.00.20.1---o------o------o------o------o------o-----(#)-09-2709-2711-0211-1511-1511-1611-16

AgentLeak 0.20.1: Security release

Released

  • A configuration file can no longer send an arbitrary environment variable to a host it chooses: a provider's key goes only to that provider, another *_API_KEY only to a local server, anything else only when the operator allows it in AGENTLEAK_KEY_ENVS.
  • Closes the case where a pull request edits the CI configuration to send a secret the job exposes to its own host. Setups using Azure OpenAI or a company gateway need one line in the environment; the run prints it.
  • No new feature and no score change.

AgentLeak 0.20.0: Two new packs, AI-security mapping, finding triage, imports from other tools

Released

  • Adds packs derived from AgentSocialBench (105 scenarios) and MCPTox (24); the corpus grows from 362 to 491. Reports map findings to the OWASP Agentic Top 10, MITRE ATLAS and NIST AI 100-2 terms.
  • Adds finding triage in the web workspace, agentleak import for promptfoo, Inspect AI, garak and PyRIT outputs, SARIF for traces and a pull-request comment in the GitHub Action.
  • Context rules raise recall on values written after a cue word: held-out English cued prose goes from 0.604 to 0.881. 23 of the 362 earlier scenarios change score, 6 of them verdict.

AgentLeak 0.19.1: Security release

Released

  • Closes two critical issues in web deployments: a request could have the server send an environment variable to a host it chose, and a project's model URL could receive the operator's key. Operators of 0.19.0 or earlier should rotate secrets held in the server environment.
  • Fixes 19 further issues: DNS rebinding past the outbound guard, MCP tools reading outside the project, quadratic analysis and code-scan costs, uncapped chunked bodies, and others listed in the changelog.
  • No new feature and no score change: the full benchmark gives byte-identical rows. On the hosted instance, the LLM judge no longer runs for tenant requests.

AgentLeak 0.19.0: Research workflow, agent self-test, wider detection, security fixes

Released

  • Adds agentleak bench (manifests, paired statistics, LaTeX export, reproducibility capsules) and agentleak selftest with nine MCP tools.
  • Fixes cross-account scenario access, unescaped trace content in the HTML report and unmasked values in code-scan snippets; hosted deployments refuse model endpoints on private addresses.
  • Adds check-digit validated identifiers, provider token rules and an InjecAgent-derived pack. 27 scenarios change verdict and the corpus grows from 266 to 362.

AgentLeak 0.18.0: Spanish identifiers, quieter Presidio tier, faster analysis

Released

  • Adds validated Spanish DNI/NIE and social security numbers, Spanish dates of birth, addresses and health terms.
  • Fixes Presidio-tier false positives on short numbers and a silent Presidio load failure; reports a dropped tier as degraded.
  • Indexes the obscured-copy search: about 35 times faster on the 1,000-event high-finding benchmark, with identical findings.

AgentLeak 0.17.0: Model gateway and evaluation integrations

Released

  • Adds an OpenAI-compatible LLM proxy and a LiteLLM guardrail.
  • Adds a promptfoo assertion and optional task-success reporting alongside privacy scores.

AgentLeak 0.16.0: Response inspection and traceable policy decisions

Released

  • Adds opt-in tool-response inspection, tokenization, flow-rule groups and origin tracking.
  • Adds signed evidence logs and MCP tool pinning; improves prose detection.

AgentLeak 0.15.0: Encoded disclosures between agents

Released

  • Detects supported encoded, obscured and fragmented copies of sensitive values across agent messages.
  • Extends redaction and gateway handling to supported encoded values.

AgentLeak 0.14.1: Scenario-library count correction

Released

  • Distinguishes loaded scenarios from the 266 scenarios available in the bundled library.

Selected recent releases. Read the complete software changelog.