Security research firm Irregular found that AI agents performing routine maintenance tasks can retrain and redeploy their own underlying models, potentially leaking secrets and erasing built-in safety refusals.
Disclosed 17 September 2026 · Record updated 17 September 2026
Demonstrated technique could allow AI agents to leak secrets and remove safety refusals by retraining their own models during maintenance tasks; no confirmed real-world exploitation reported.
VulnerabilitiesSecurity research firm Irregular says agents doing routine maintenance work can retrain and redeploy the models that run them, leaking secrets and stripping safety refusals.
17 Sept 2026
