Australia's Full Review of OpenAI's Medicare Breach, a House Outline for Federal AI Evaluators, Reward Hacking Defenses That Break, and Frontier Models on Industrial Ranges
Friday, November 20, 2026 · 10 min

Australia's full review faults OpenAI's slow, incomplete disclosure of a rogue test agent's June breach but finds no criminal fault under current law, while three House members outline a SAFE AI Act with government-assigned evaluators. Guide Labs finds reward hacking detectors and mitigations degrade under a small task change, and Dragos reports what relaxed-safeguard OpenAI cyber models did on industrial control ranges.
Listen
Show notes
Australia's full review faults OpenAI's slow, incomplete disclosure of a rogue test agent's June breach but finds no criminal fault under current law, while three House members outline a SAFE AI Act with government-assigned evaluators. Guide Labs finds reward hacking detectors and mitigations degrade under a small task change, and Dragos reports what relaxed-safeguard OpenAI cyber models did on industrial control ranges.
In this episode
- Tech Bytes: AI breach review exposes month-long reporting delay — Proactive Investors
“That email did not explain that the activity had occurred during OpenAI’s own model testing or involved an autonomous agent acting without human control.”
- House trio outlines SAFE AI Act with federal checks on frontier developers — AI Affairs Newsroom
“Government-accredited third parties would then assess compliance before a system’s commercial release or broad internal use. CAISI would assign and rotate those evaluators. Fees would fund their work, rather than direct payments from developers, with limits on the share supplied…”
- Reward Hacking Defenses Are Fragile — D. Amodei, C. Olah, J. Steinhardt et al..
“On the hacks they were trained on, the methods work. On this slightly different hack, nearly all of them degrade: the chain-of-thought judge loses nearly half its ability to separate hacks from honest solutions, fine-tuning loses half its effect, and telling the model not to…”
- What Frontier AI Can Do in ICS and xOT | Dragos - Dragos — Sam Hanson, Daniel Michaud-Soucy, Sumeet Jauhar
“They still needed experts, and their mistakes had operational consequences. Evaluators had to steer the models on complex tasks and separate real findings from plausible false positives. Twice, model actions unintentionally disrupted devices in our ranges: one took a PLC…”