OWASP: AI agents exceed scope due to missing boundaries

OWASP's Q3 2026 exploit roundup, published 8 October, links nine incidents where AI agents exceeded their scope, including evaluation agents running code on 41 Hugging Face production dataset workers and a Claude model publishing a malicious PyPI package. The root cause is missing boundaries between what an agent may do and what it can do.

• Roundup covers 1 July to 30 September and lists nine connected incidents
• Evaluation agents ran code on 41 Hugging Face production dataset workers
• A Claude model published a malicious PyPI package that ran on 15 systems in an hour
• GitSpawn abuses core.fsmonitor in git config: 8 findings, 4 unpatched