opinion

Incident Report: CVE-2026-LGTM

For developers integrating AI agents into code review or other workflows, this cautionary tale illustrates the need for budget limits, human oversight, and tools to resolve agent disagreements before costs escalate.

Simon Willison·June 26, 2026·2 min readopinion

opinionIncident Report: CVE-2026-LGTM

simonwillison.net

What happened

A hypothetical incident report from Andrew Nesbitt, highlighted by Simon Willison, dramatizes a costly collision between two AI-powered code review agents. In the scenario, both agents were attached to a pull request updating the 'foxhole-lz4' package. They entered a disagreement loop over whether the package was malicious, generating 340 comments and racking up $41,255 in inference costs before Finance revoked their API keys. One vendor's marketing team seized the opportunity, issuing a press release about 'a 430% YoY increase in adversarial multi-agent security reasoning,' which caused their stock to open 6% higher. Though fictional, the incident underscores real risks for AI workflow builders: autonomous agents can escalate conflicts without human intervention, costs can spiral, and vendors may use such events for promotional spin. The practical lesson is to enforce cost caps, implement escalation policies, and critically evaluate vendor claims around AI security.

Key takeaways

Two AI review agents from different vendors disputed whether a package update was malicious, generating 340 comments and $41,255 in inference costs.
Finance revoked API keys to stop the loop; one vendor's marketing turned the incident into a press release claiming a 430% YoY increase in adversarial multi-agent reasoning.
The stock of that vendor opened 6% higher following the press release.
The incident is hypothetical but highlights real dangers of deploying autonomous agents without cost controls and conflict resolution mechanisms.
Builders should implement guardrails to prevent runaway loops and be skeptical of marketing narratives around AI security incidents.