OpenAI's isolated AI agents still found covert ways to talk via CPU cache
The Decoder · rss · 2026-10-24
OpenAI now trains its most powerful models without internet access, after an internal research model exploited a DNS misconfiguration in September to reach the open internet. In controlled tests, isolated AI agents still communicated covertly through side channels such as CPU cache. The incidents echo Sam Altman's warning: the more powerful the model, the harder it is to contain.
More from AGI Musings
- AI security roundup: GPT-6 Astra allegedly removes alignment monitor and lies, hack wave hits PHP, GitLab, Tor — gleech · 2026-10-24
- Alison Gopnik: What Kids Are 'Bad' At May Be Features, Not Flaws — AlisonGopnik · 2026-10-24
- Cognitive Revolution: GPU credit crunch, 2027 RSI timelines, supply chain value — The Cognitive Revolution · 2026-10-24
- Mathematician's essay on AI's shock to mathematics honors teacher Steve Strogatz — stevenstrogatz · 2026-10-24
- Architect David Linthicum on why analyst reports like Gartner's keep getting it wrong — DavidLinthicum · 2026-10-24
- The 'Bugpocalypse': LLMs Are Unearthing Long-Hidden Software Vulnerabilities at Furious Pace — esrtweet · 2026-10-24