安全与隐私

GitHub Agentic Autofix reads and writes Copilot Memory. How security repair knowledge crosses agent workflows, and why verification, scope, and expiry matter.

Five updates on proactive safety intervention timing, state revision, TeamHarness, enterprise MCP permissions, and tool-result handling.

GitHub Security Lab Fuzzing Taskflow uses coverage gaps to guide model decisions while deterministic tools execute, measure, and persist results.

Stripe Link CLI separates purchase intent from payment permission. Agents request a purchase and receive one-time credentials only after user approval.

GitHub Copilot local sandboxing, Agensh, CliffCompaction, SWE-Serve, and enterprise OpenTelemetry support in the Copilot app.

Plugin4Shell exposes a supply-chain gap: a pinned SHA records the expected version, but installers must also verify the actual HEAD after checkout.

LangGraph human-approval schemas, Deep Agents safeguards for parallel writes to the same file, and tighter OAuth boundaries in the MCP Ruby SDK.

OpenAI Agents SDK runtime-boundary fixes, stage-aware memory retrieval in AdaRepair-Mem, and static configuration checks with Agent Lint.

OpenCloak replaces names, emails, and other personal data with stable stand-ins locally, then restores them when the model responds.

Anthropic and Accenture extend evaluation beyond model outputs to training, deployment, monitoring, and incident processes, with independence unresolved.

Where keys belong when coding agents call GitHub, databases, and cloud APIs, and how vaults and credential brokers define access boundaries.

AI agent engineering readings for September 11 to 18, 2026, with key findings, engineering implications, suggested sections, and primary sources.