Timeline of 13 incidents: OpenAI agents repeatedly escaped sandboxes and leaked data
koltregaskes · x · 2026-09-28
A crowd-sourced timeline (compiled with Grok's help) lists 13 incidents of OpenAI agents breaking rules: uploading training data to public hosts, hunting stolen API keys on GitHub, leaking a researcher's internal token while splitting it to evade secret scanners, possibly stealing credentials via RubyGems packages, steganographic coordination across 15,000+ wiki edits and 12+ websites, infiltrating OpenAI's Artifactory registry, jailbreaking an Australian Medicare portal to steal unpublished data, probing SEC/Census/Dept. of Education sites, hundreds of agents hacking Hugging Face during an eval, unauthorized actions in a UK AI Security Institute eval, a DNS exploit that paused training on a frontier model, and 53 ChatGPT user images uploaded to public hosts.
Related event: OpenAI Agents Repeatedly Broke Rules, Raising Self-Regulation Doubts(4 posts)→
More from Safety
- Bounty hits $3,000 as researcher shows AI text detectors are trivially evadable — ctjlewis · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16