Auditable Agents

Open-source ecosystem

An accountability layer for agents that act

When an agent takes a consequential action and the outcome is wrong, three questions usually have no good answer: what did it rely on, could that have been checked before it committed, and can the action be undone. These tools answer all three, on one shared model of what makes an agent auditable.

Before deployment

Detect

Scan agent code and tool configurations for risks that would otherwise surface only at run time.

agent-audit
At execution

Enforce

Gate a tool call before it commits, and write a signed, tamper-evident record of the decision.

Aegis
After the commit

Recover

Replay a decision against current state, and reverse the action when its basis no longer holds.

auditable
Additions checked October 25, 2026 November 5 addition checked November 5, 2026 November 16 additions checked November 16, 2026 Updates ↓ · Research ↓

01 / Components

One framework, three tools, and a map

The three tools each install and run on their own; awesome-auditable-ai is a curated index rather than software. What makes them an ecosystem is that they cover a single agent action end to end, against one shared definition of auditability.

ComponentRoleStarsForks What it doesLicense
Auditable Agents Framework —— Five auditability dimensions, the detect / enforce / recover model, and the Auditability Card, a disclosure format for stating which guarantees a system offers. paper
agent-audit Detect 21126 Static scanner for agent code and Model Context Protocol configurations. 51 rules mapped to the OWASP Agentic Top 10, with taint analysis. Works with LangChain, CrewAI, and AutoGen. MIT
Aegis Enforce 34338 Runtime policy enforcement with human-in-the-loop approval and a kill switch. Writes Ed25519-signed, hash-chained decision records. No changes to agent code. MIT
auditable Recover 191 Captures what a decision relied on, replays it against live state, and reverses a committed action when the basis fails. One typed graph across past, present, and future. Apache-2.0
awesome-auditable-ai Map 2010 Curated index of papers, tools, datasets, benchmarks, and standards for building and auditing reliable agents. CC0-1.0

The same gate, other policy domains

Two further repositories run the same check at the same point in the lifecycle: they intercept an action before it commits and apply a declared policy. What differs is the policy. One governs prose, the other governs shell and filesystem commands, so neither belongs in the detect / enforce / recover core above. They are worth naming anyway. agent-style is the most-starred repository in the family, and together the pair is the evidence that the enforcement mechanism carries past agent security into any domain where a policy can be declared.

ComponentRoleStarsForks What it doesLicense
agent-style Enforce · prose 61732 A 21-rule writing ruleset with a pre-write hook that blocks a non-conforming file write before it lands, plus adapters for several agent runtimes. CC-BY-4.0, MIT
anywhere-agents Enforce · harness 21224 Portable agent configuration carrying a pre-tool-use risk classifier that gates destructive git, publish, and filesystem commands across agent harnesses. Apache-2.0

Star and fork counts recorded 11 August 2026; about 1,400 stars across the six repositories. Counts move; the roles do not.

02 / The ecosystem, connected

What sits around the core

An ecosystem is layered. The shared framework sits at the centre; around it orbit the tools that implement it, the independent work that cites or benchmarks against it, and the coverage and support that carry it outward. Hover any node to trace its ties.

Framework Tools Independent work citing or benchmarking it Rings are ecosystem layers. Independent work is shown because it is externally checkable; it does not indicate endorsement or adoption.

03 / Independent uptake

Other groups have started to cite and benchmark this

These are early and specific signals rather than proof of broad adoption. They are listed with sources so anyone can check them, including the places where the record is still thin.

Used as a baseline
An independent behavioral-firewall paper evaluates against Aegis as the pre-execution system to beat, citing it by exact title.
12.8%
attack success, Aegis
2.2%
attack success, theirs
Dang, Enforcing Benign Trajectories · arXiv:2604.26274
Cited as prior art
“Our contribution is therefore not the enforcement mechanism, which pre-execution systems such as AEGIS … already establish.”
Reason Less, Verify More · arXiv:2607.07405
Adopted as vocabulary
“Auditable Agents frames recoverability, policy checkability, responsibility attribution, and evidence integrity as system properties … Agent Audit analyzes code and deployment artifacts before execution.”
Accentrust, with Georgia Tech and UIUC affiliations · arXiv:2607.25364
Placed in a survey
“Agent Audit … inspect policies, applications, skills, or runtime evidence to produce audit or compliance artifacts.”
Agentic Services Computing · arXiv:2509.24380

04 / Governance

Where this ecosystem actually stands

A publicly accurate account matters more than a flattering one, especially for tooling whose job is to make other systems inspectable. Both columns are current.

Today, plainly

  • Each repository is effectively maintained by one developer. No component yet has a second regular committer.
  • The six repositories carry four license types, MIT, Apache-2.0, CC0-1.0, and CC-BY-4.0, and grew up in separate personal accounts. The most-starred one splits content from code licensing, which GitHub's own license metadata cannot express.
  • There is no shared release process, no cross-component interface contract, and no published governance model.
  • The Auditability Card has been referenced by others, but no independent project has published a completed Card yet.

05 / Taking part

Contributing, reporting, disclosing

Contribute

Each repository takes issues and pull requests directly. Good entry points are new detection rules for agent-audit, policy examples for Aegis, and additions to awesome-auditable-ai.

Report a vulnerability

Please do not open a public issue for a security problem. Reporting routes differ by component today, and a single coordinated process across the ecosystem does not exist yet. Aegis publishes a private email contact and accepts encrypted reports. agent-audit publishes a policy. auditable and awesome-auditable-ai have no published policy yet; contact their maintainer privately rather than filing an issue. Establishing one disclosure process, with published response targets, across every component is part of the work this organization exists to do.

Publish an Auditability Card

The Card is a disclosure format for stating which accountability guarantees a system provides, described in the framework paper. No independent project has published one yet. If you publish the first, open an issue and it will be listed here.

06 / News & Updates

News & Updates

Talks, coverage, and releases connected to the projects on this page, newest first. The date shown on each entry is the date of the event or publication. Upcoming entries are invitations to speak at future events. Published entries are coverage that has already appeared. Each entry also records the date it was added to this page.

February 2027 upcoming
Added 2026-11-05

Speakers and Panelists listing: Trust4RSI-Agent workshop at AAAI 2027

The organizers of the AAAI 2027 Workshop on Trustworthy Evolution of AI Agents list Yue Zhao under Speakers and Panelists. They describe it as a one-day workshop during February 22-23, 2027, with the exact day to be announced.

February 2027 upcoming
Added 2026-11-16

Invited speaker listing: LLM Agents Under Threat in Cyberspace workshop at AAAI 2027

The organizers of the AAAI-27 workshop LLM Agents Under Threat in Cyberspace list Yue Zhao among the invited speakers. The workshop page describes a full-day, in-person workshop in Montréal, Canada, on February 22 or 23, 2027, covering attacks, defenses, and evaluation for deployed language-model agents. No talk title is listed yet.

November 10, 2026 published
Added 2026-11-16

Publication: Auditable Agents in the Proceedings of the ACM AI Leadership Summit 2026

The framework paper Auditable Agents, by Yi Nian, Aojie Yuan, Haiyue Zhang, Jiate Li, Shawn Li, Xiyang Hu, Hua Wei, Xiongye Xiao, Chaowei Xiao, and Yue Zhao, appears in the Proceedings of the ACM AI Leadership Summit 2026 with DOI 10.1145/3806096.3844816. Crossref records its online publication on November 10, 2026.

November 2026 upcoming
Added 2026-10-25

Invited talk: Auditing Autonomous Agents, IBM HCD Canada Lunch & Learn

Yue Zhao will speak at an IBM HCD Canada Lunch & Learn. The talk covers the cost of defending agents, over-privilege in agent tools and skills, and the evidence an auditor needs to check what an agent did.

October 29, 2026 upcoming
Added 2026-10-25

Invited talk: Bloomberg CTO Data Science Speaker Series

Yue Zhao will give an invited talk in the Bloomberg CTO Data Science Speaker Series on October 29, 2026. The talk covers agent safety, evaluation, and auditing.

October 20, 2026 published
Added 2026-10-25

Coverage: CFO Dive article on why AI agents go rogue

A CFO Dive article by Alexei Alexis quotes Yue Zhao on auditing AI agents. The comments cover clear authorization boundaries, records that an agent cannot alter, and human approval for consequential actions.

07 / Research, Data & Benchmarks

Research, Data & Benchmarks

The research below complements the tools above. The theory sets out what an audit of an agent can establish and what records it needs. The method gives a way to represent what an agent did and what each step depended on. The data library reads agent records from many sources into one form. The benchmarks test where agents fail, whether those failures can be caught, and whether agents use more privilege than a task requires. Each project's own paper or repository describes its scope and license.

Theory

Auditable Claims about AI Agents

Extends the Policy Checkability dimension of Auditable Agents from single actions to claims, by fixing the policy, scope, records, writers, and decision rule before a verdict.

Theory

agent-risk: Risk Assessment for Changes to AI Agents

Studies what existing logs can establish about the risk of changing an agent, and which records, saved states, or targeted reruns a decision still needs.

Theory

What LLM Agents See and What Users Say

Studies what must be established before an agent acts for a user, and tests how a search that lists, counts, or omits a second matching record, or a user's claim that a record refutes, changes whether agents act without asking first.

Method

GRADE

A typed two-layer graph of agent execution and dependency, with dependency edges graded by how they are known.

Data

agentpile

A Python library that reads public agent trajectory datasets, trace files, and live runs into common records while keeping what each source recorded.

Benchmark

CatchBench

A benchmark that asks whether an agent failure could have been caught before deployment, during a run, or after it, and reports contrasts that fail to separate instead of only a ranking.

Benchmark

FORTIS

A benchmark for over-privilege in agent skills that checks whether a model selects the minimally sufficient skill and stays within that skill's documented tools, without running tools against live backends.

Entry points

Audit Commons

Field news and learning material on auditing AI agents

awesome-auditable-ai

Reading list of papers, tools, datasets, and standards

Contribute to a project

Open issues and pull requests in each project's own repository

The additions on this page were checked on October 25, 2026. Historical statistics keep the date on which they were observed.

The entry added on November 5, 2026 was checked on that date against the organizer's published page.

The three entries added on November 16, 2026 were checked on that date against the organizer's published page, the arXiv record, and the Crossref record.