Verify the source.
Test the result.

An extension update is not necessarily a complete incident response.

Cursor’s October 30 update and Microsoft’s investigation describe tampered extension copies, versions 2.1.312 and 2.1.313, distributed through Cursor’s marketplace for Windows and macOS. The reported exposure window was October 29, 02:27–13:35 PDT. Persistence outside the extension means uninstalling or updating it alone can be insufficient.

Evidence & limits. This is an October 30 incident update, not a claim that the incident began that day. The notices do not establish impact to the official VS Code Marketplace or npm distribution, or compromise of Anthropic’s systems. Use the vendors’ current remediation instructions rather than assuming every similarly named package is affected.

Inspect the enforcement configuration, not just the presence of a proxy.

October 30 commits expose a small API and repair server-sent-event streaming. The existing proxy sits before a harness executes tool calls, applying undeclared-tool and host checks, detectors with severity levels, and an optional judge. This is an integration update to a project already present October 29, not a new launch.

Evidence & limits. The quickstart’s log-only mode does not block. Judge errors allow and log by default unless fail-closed behavior is configured. This is not a sandbox or a guarantee against unsafe execution; effectiveness depends on coverage, configuration and the surrounding harness.

Savings on jointly solved tasks do not establish savings on every task.

The LFM2.5-24B-A2B adapter trains with AsyncGRPO on 1,000 tasks, using correctness rewards and rank-32 LoRA. Its author evaluates 100 test tasks with one attempt at temperature 0.8. Reported pass@1 falls 4.2 percentage points; tool calls fall 39% only on the subset both versions solve.

Evidence & limits. The author characterizes the accuracy change as within ±5-point noise, not a statistical demonstration of equivalence. Checkpoint 100 was selected after accuracy dipped at 110. The GGUF conversion was not re-evaluated. Small, author-run results do not establish general savings or unchanged reliability.

Benign-looking data can pass a content monitor yet produce unwanted downstream behavior.

Parallel research agents use a shared log to design datasets of at most 5,000 examples, fine-tune a separate research model, and iterate against monitors, behavioral evaluations and cheat detection. Fiction and roleplay examples can look acceptable individually while inducing misbehavior after training.

Evidence & limits. These are author-reported experiments, not independent replication. Search agents had full monitor access and unconstrained data choices; their examples formed the entire training corpus. The resulting models were never deployed. The official announcement falls on October 30 in New York; the report itself is month-dated October 2026.

A registry becomes a control when unlisted systems cannot be used.

The consultation proposes a board-approved AI policy, risk-tiered approvals, and a registry of active, inactive and retired systems. New or changed deployments would undergo validation and impact assessment. A senior-executive, cross-functional committee would oversee the program, with supplier accountability and records covering tests and incidents.

Evidence & limits. These are proposed guidelines, not enacted requirements or measured implementation results. Comments are due November 27. The official page is dated October 30; an exact publication timestamp was not established, so its New York-day boundary is uncertain.