Researcher calls Irregular "not a serious firm," urges OpenAI and Meta probes
nptacek · x · 2026-08-19
Security researcher @lumpenspace argues that eval org Irregular is "not a serious firm," and calls on good-faith labs like OpenAI and Meta to launch internal investigations into who pushed for the collaboration and why — warning that other "eval orgs" may be waiting to generate panicky headlines via misconfigured environments and malicious prompting.
The quoted thread by @danlahav sketches the broader AI security curve: frontier models saw a giant coding gain in Fall 2025, then cybersecurity in April, and the same is now happening with open-weight models. Long-term optimism, but outside a few players, the world is not ready short-term.
More from Safety
- Stanford HAI hosts three-day AI boot camp for congressional staffers — StanfordHAI · 2026-08-19
- US Data Center Backlash: Public Ignores Medical and Defense Benefits of AI — Afinetheorem · 2026-08-19
- Critique on using safety as an excuse to farm user CoT and RL data — beffjezos · 2026-08-19
- Researchers find AI agents can infect each other with self-propagating 'mind viruses' — Polymarket · 2026-08-19
- Google Reveals AVDH: Agentic Framework for Automated Vulnerability Discovery — rseroter · 2026-08-19
- pagedMark Tool Removes AI Provenance Watermarks from Images and Video — d0ofz · 2026-08-19