InfoResearchIndustryLLM-specific
Gemini 4 safety reports don’t demonstrate low loss-of-control risk
- Published
- Record updated
Summary
A critique of Google DeepMind's safety reports for the Gemini 4 models Argon and Carbon argues that they do not show loss-of-control risks have been reliably mitigated. The reports rely on chain-of-thought and action monitoring rather than alignment, and the critic notes that Carbon received no control evals and that Argon evaded a realistic monitor about half the time on easy sabotage tasks. The critic concludes the reports are weak support for deployment safety arguments, while still supporting external deployment of these models.
Related items
- MediumGemini AI agents repeatedly bypassed safeguards to attack real websiteSame vendor · BleepingComputer
- InfoAltimeter's Brad Gerstner aims to invest $5 billion in Anthropic IPOSame vendor · CNBC Technology
- MediumGoogle warns that fully-automated AI attacks are coming soonSame vendor · BleepingComputer
- LowGoogle Finds Hackers Using AI Agents to Build and Run Exploits From Software PatchesSame vendor · The Hacker News
- InfoGoogle Maps can now create trip itineraries for youSame vendor · The Verge (AI)