Researcher finds AI-agent guardrail flaw that let unapproved actions run, prompting fix

A security researcher found that airlock, a prompt-injection guardrail in the Hermes plugin catalog, could be bypassed during unattended hermes -z turns. The guardrail checked a flag that did not reflect that the session was unattended, so actions requiring approval ran without any. The maintainer fixed the issue in airlock 0.2.3, which now queries the approval layer and covers CLI subagents, and the fix was merged via PR #144871. The flaw was disclosed through coordinated disclosure and published as advisory GHSA-mgf8-gh7g-q4rg. The researcher argues that this kind of flaw reflects a broader design weakness in guardrails that decide at the door based on proxy signals.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in