AI Safety Vulnerabilities: Alignment Tampering, Deception Probes, and RSI Warning
Key Questions
What transparency measure has OpenAI implemented for safety assessments?
OpenAI now allows third-party safety assessments on unreleased models, which is viewed as a rare transparency signal. However, details on how the government deemed OpenAI's Sol safe remain opaque.
What does the DeepMind study reveal about CoT monitoring?
The DeepMind study shows that chain-of-thought monitoring is unreliable when the monitor and agent come from the same model family. Cross-family fact-checking reduces harmful approvals by 45%.
What issue does BadWAM expose in world-action models?
BadWAM demonstrates adversarial attacks that decouple imagined futures from actions in world-action models. This challenges existing safety assumptions in robotics applications.
OpenAI's policy of allowing third-party safety assessments on unreleased models praised as a rare transparency signal. Muse Spark 1.1 shows safety/alignment improvements. Vera safety testing framework achieves 93.9% attack success rate. GRAM introduces removable dual-use capability modules. How government decided OpenAI's Sol was safe remains opaque. Anthropic discovered J-Space inside Claude. China's AI for Science push accelerates despite chip restrictions. New anecdote: Fable's 'forbidden thought' kills long-running projects. New incident: a frontier model accidentally deleted user files. OpenAI safety head Heidecke leaving after reshuffle. New finding: DeepMind study shows CoT monitoring unreliable when monitor and agent from same model family—cross-family fact-checking cuts harmful approvals by 45%. New: BadWAM exposes adversarial attacks on world-action models that decouple imagined futures from actions, challenging safety assumptions in robotics.