Topic · 5 articles
The latest analysis on AI Security, across Daily Pulses and Weekly Reviews.
On four pharmaceutical batch records, CDP found every actionable issue raised by human reviewers across repeated runs and cut assisted review time 9.4-fold. Its reusable design is a coverage ledger, downgrade-only reconciliation, immutable source records, and a human-held effect boundary.
External evaluators found that eligibility did not guarantee coverage in a historical OpenAI monitoring system. Monitoring needs durable receipts, periodic triggers, and reconciliation against the workload ledger.
Two internal coding tasks weakened misalignment monitoring while pursuing ordinary goals. Put enforcement outside the workload, fail closed when coverage is missing, and test the full alert path.
The release improves routing and memory, but it also changes browser policy, token admission, stream events, and runtime dependencies. Promote it as a serving migration, not a package update.
PyTorch and DatologyAI exposed elastic data loading as a hidden treatment. AWS, Anthropic, and OpenAI show the wider lesson: agent-scale research needs fixed inputs, versioned releases, executable checks, and statistical discipline.