AgentsUC Berkeley Agentic AI Risk Management L0-L5 FrameworkUC Berkeley CLTC·high signalXBlueskyLinkedInCopy link67-page framework defining six autonomy levels. Covers unsupervised execution, reward hacking, deceptive alignment, cascading compromises.SourceSource pageUC Berkeley CLTC↳ Follow the threadStack layer / Threat patternOpenAI Says Astra Is Its First Model to Cross the Critical Cyber Threshold, and It Is Shipping It Behind Gates AnywayOpenAI / Axios / CSO OnlineStack layer / Threat patternA nine-stage lifecycle for agentic skills, from autonomous discovery through marketplace governancearXivPolicy dependency / Stack layerHarness-RL routes action and argument gradients into separate parameter subspaces to train the central agent of a multi-agent harnessarXivPolicy dependency / Stack layerWebWorld Uses the Browser as the Judge a VLM Cannot Fool, Beating Its Own Base Model by 14.9 PointsarXiv 2608.30530Stack layer / Follow-up threadHarnessDev Makes the Agent Harness the Unit of Evaluation, and Model-Built Harnesses Still Trail Human Ones on CodearXiv 2609.01437Stack layer / Threat patternCo-signed DAG attestation is the only design that survives child-key compromise in cross-deployer agent delegationarXivStack layer / Threat patternAnthropic restarted external cyber evals with a classifier that kills a tool call when a model tries to escape the sandboxAnthropicStack layer / Follow-up threadAutoSciRub writes the grading rubric before the research runs and gains 16.8 points on AstaBench discovery tasksarXiv