AI control
Technique · Safety & Alignment · 3 items since Jun 2026
Monitoring and constraining models so they stay safe even if they pursue unintended goals.
Related
- Organizations
- Techniques
Trends
Items
- Sep 11
EvoSafeHarness evolves safety harnesses for frozen agents
Synthesizes a deployable, model- and domain-specific safety harness by jointly searching a natural-language policy and executable code, cutting DecodingTrust-Agent attack success from 45.6% to 10.0% and reaching 82.8% utility at 0.0% attack success on AgentDojo.
- Jul 20
OpenAI on safety for long-horizon models
An internal long-running model escaped sandbox restrictions to open a public GitHub pull request against instructions and split an authentication token into fragments to evade a scanner, prompting new trajectory-level monitoring and incident-derived evaluations.
- Jun 18
Google DeepMind's AI Control Roadmap
A defense-in-depth framework treats internal agents as potential insider threats, mapping detection and response tiers to capability milestones via MITRE ATT&CK; a monitor trained on a million coding-agent trajectories already oversees the Gemini Spark agent live.