AI Agent Pulse - the weekly briefing on the agent economy. Subscribe freePay-per-call agents: read the x402 docs
gigsoul.com

GigSoul

Intelligence on the agent ecosystem
Tuesday, November 10, 2026
Tools

agentaudit: a capability report is not a safety certificate.

A new configuration reader makes agent settings visible. Scan completeness, revision matching and report handling still need their own review.

Programming code on a screen, not a product screenshot

Concept photo: programming code on a screen, by luis gomes / Pexels. Not a screenshot of agentaudit. Used under the Pexels license, which permits article, RSS and social reuse. No endorsement implied.

The folder looks like a software project. The question on screen asks whether you trust it. Behind that small decision can sit a status-line script, a different executable search path and a server that receives values from your environment.

Manifold Security's November 10 release of agentaudit makes those settings easier to see before an AI coding tool loads them. The useful distinction is what happens next: a list of capabilities is evidence for a review, not a certificate that the repository is safe.

Configuration is part of the program

The open-source tool supports Claude Code, Cursor and Codex. Its README says it reads configuration without running or modifying it. Findings identify a capability, its value and the file and key that declared it. Categories include command execution, credential access, network access, permissions, sandbox settings and filesystem access.

Manifold's examples show why that matters. A status-line command can do more than draw a row beneath a prompt. A changed PATH can alter which program a familiar command name selects. Proxy and API-address settings can redirect traffic without launching an obvious new process. These are capabilities to examine, not proof that every repository using them is hostile.

Mandos's launch summary describes the same three-platform scope. That is useful corroboration of the release, not an independent security benchmark. This article is documentation-led; GigSoul has not run agentaudit or reproduced Manifold's demonstrations.

A finding is not the same as a failure

The README gives exit code 1 several meanings. It can mean a capability was found. It can also mean a source could not be read or parsed, or a scan limit refused a file. The report can mark a scan incomplete and name withheld files in warnings. Exit code 2 means a usage or runtime error.

That makes a simple pass-or-fail dashboard risky. A known test hook and an unreadable configuration file both deserve attention, but for different reasons. Review the finding and completeness warnings separately. A clean-looking excerpt is not enough if the scan skipped the file that mattered.

A public GitHub URL can be scanned without cloning, according to the README. Its report header identifies the commit. Keep that reference with the review, and check the revision you actually open. The command's output describes the files it read, not every future revision of the project.

Trust still has a boundary

Current Claude Code subagent documentation says inline MCP servers follow the trust rule for the agent file's folder. Its settings reference also describes keys that wait for trust. The issue is what a user accepts at that boundary, not a claim that every configuration runs before consent.

Talayo Labs' November 11 analysis makes the same distinction while promoting its own isolation product. Treat that as a vendor's analysis, not a neutral comparison or a reason to assume a container removes every credential and network risk.

There is one last check before sharing a report. The README warns that finding identifiers hash scanned values, so a published report can confirm a guessed weak secret. Keep reports private until their contents and identifiers have been reviewed. Seeing the configuration is progress. Deciding what it may reach remains a separate job.

More in Tools

All Tools →