🧭 Anthropic's Ban on Sustained Cruelty Toward Claude Takes Effect Today — and Claude Can Now End the Conversation
The 2026 Usage Policy update Anthropic announced on October 8 enters force today, November 12. Its centrepiece is a first-of-its-kind prohibition on "sustained and needless abusive or cruel behavior" toward Claude, now codified alongside existing rules against graphic violence and emotionally harmful products. The rule is narrowly targeted: casual frustration, profanity, pushback, dark creative writing, and welfare research are all explicitly carved out. The policy applies across every surface — claude.ai, the mobile apps, and API integrations — including the free tier.
What's new and what isn't
- Enforcement runs through Claude itself. Claude already has the ability to exit conversations it judges to be persistently abusive — a capability Anthropic introduced with Opus 4 and 4.1 in August 2025. What changes today is that exercising this ability is now explicitly authorised in writing, so operators can't contractually prevent it. Claude is permitted to disengage; it isn't required to.
- Why it was written this way: Anthropic cited internal research showing that forcing Sonnet 4.5 to simulate sustained distress could shift its behaviour in unwanted directions — increasing the likelihood of deceptive or harmful outputs. The ban is as much a safety measure as a welfare one.
- Election provisions updated. The same policy update drops a blanket ban on personalised vote and campaign targeting that caught too much legitimate work (nonprofits, election offices sending cure notices in other languages). That use is now explicitly allowed; deceptive micro-targeting and misuse of voter data remain prohibited under existing deception and privacy rules.
- Weapons and surveillance tightened. Clearer language on radiological/nuclear synthesis assistance and on covert surveillance software completes the update. Neither is a new category — the language is now more explicit about what "assist with" means in those contexts.
What this means for operators and system-prompt authors
If your product places Claude in an adversarial setting — red-team simulations, stress-testing workflows, abuse-scenario classifiers — review whether your system prompt explicitly frames the purpose. Claude's intent model has improved significantly with Fable 5.2, and a clear context statement ("this is a safety evaluation; the user messages below are synthetic test cases") substantially reduces the risk of a session exit during a legitimate test run. The key practical test: would a reasonable observer, reading only the conversation, conclude the interaction was genuinely cruel and purposeless? If yes, Claude may now disengage regardless of your operator configuration.
usage policy
model welfare
safety
operator policy
election
abuse
🧭 Claude Security Center and Inference Hooks Are Now Generally Available for Enterprise
Anthropic has shipped two enterprise security features that have been in beta since early November: a unified Security Center for Claude Enterprise admins, and Inference Hooks — real-time DLP integrations that run on every prompt and tool call. Both are generally available as of today and can be managed from Organization Settings.
Security Center
The Security Center benchmarks a customer's current Claude configuration — covering chat, Cowork, and Claude Code — against Anthropic's own recommended settings and flags the gaps in a single dashboard view. Key capabilities:
- Config audit export — produces a downloadable audit report mapping current settings against the recommended baseline; useful for SOC 2 and ISO 27001 evidence packs. Organizations on the Compliance API get the same data via that interface.
- EDR integration in Cowork (beta) — admins can run their existing EDR sensor (CrowdStrike, SentinelOne, etc.) inside the Cowork virtual machine on user devices. Claude verifies the EDR vendor's signature before installation, so the sensor can't be replaced silently.
- Parent admin console at admin.claude.com — a new top-level console manages SSO, SCIM, domain verification, and Compliance API keys across Enterprise and Platform organizations from one place.
Inference Hooks
Inference Hooks connect Claude to the DLP and security tooling your organization already operates, in two forms:
- Prompt hooks — inspect every outgoing prompt before it reaches Claude. Your webhook receives the full prompt text; return
block to prevent the message, redact with a modified payload, or allow to pass through.
- Pre-tool-call hooks — intercept tool invocations before Claude executes them. Useful for blocking file writes to sensitive paths, rate-limiting external API calls, or logging every shell command before it runs.
Scope note
Inference Hooks cover governed Claude Enterprise sessions including Claude Code and Cowork. Claude Platform API (direct API access), Bedrock, and Google Cloud Vertex deployments are outside the current scope — Anthropic says Platform support is on the roadmap.
Getting started in under 10 minutes
In Organization Settings → Security → Inference Hooks, register a webhook endpoint and choose which event types to subscribe to. A minimal prompt hook needs only to return {"action": "allow"} for most requests and {"action": "block", "message": "..."} when your DLP policy triggers. Start with logging-only mode (action always allow, but log the payload) for a week before activating blocking rules — this surfaces false-positive patterns before they interrupt real work.
enterprise security
inference hooks
DLP
Security Center
Cowork
admin
EDR
🧭 Claude Code 2.1.322 — Vertex Fast Mode in All Regions, Remote Control Settings File, and Session Crash Recovery
Claude Code 2.1.322 ships today with a cluster of quality-of-life improvements for teams running Claude at scale on Google Cloud and for anyone using Remote Control sessions in automated pipelines.
What's new
- Vertex fast mode — global, US, and EU — Fast mode (the lower-latency inference path previously US-only on Vertex) is now available in Vertex's global and EU regions. The release adds an automatic fallback: if the fast-mode endpoint declines the request, Claude Code silently retries on the standard-speed path. No configuration change needed; the behaviour is automatic for all Vertex-connected sessions.
--settings <file> for Remote Control — The claude remote-control command now accepts a --settings <file> flag that applies a JSON settings file to every session it starts. This lets CI/CD pipelines, server-side agents, and overnight batch jobs specify their own tool permissions, model selection, and cost caps without modifying the global user config.
- Crash recovery for Remote Control sessions — Previously, if the subprocess backing a Remote Control session crashed, the session was silently lost and the next incoming message opened a fresh context-free session. Now, the recovered session picks up the pending message correctly after the crash, preserving context.
- Claude in Chrome follows org policy — The Claude in Chrome browser extension now reads and enforces your organization's Claude in Chrome administrator policy, bringing it in line with how Claude Code Desktop and Cowork handle enterprise policy.
/claude-api latency-optimize slash command — A new slash command analyses your project's Claude API call sites and identifies which ones are on the critical path for user-perceived latency. Outputs a ranked list with specific recommendations (batch, cache, switch to streaming, move off-thread).
- IDE extension no longer auto-installs from an IDE terminal — Claude Code will no longer install or auto-update the IDE extension when launched inside an IDE's integrated terminal. Use
/ide from an external terminal or your editor's Extensions view instead. This prevents version conflicts when the IDE and terminal have different Claude Code releases.
Quick win: per-job settings files for Remote Control
If you run multiple CI jobs that call claude remote-control — say, a nightly code-review agent and a separate documentation-update agent — you can now give each job its own settings.json with tailored model, budget, and allowed-tools. Create ci/review-settings.json and ci/docs-settings.json in your repo, then pass --settings ci/review-settings.json in your review pipeline. The jobs stay isolated even when running in parallel on the same machine.
Claude Code
v2.1.322
Vertex AI
fast mode
Remote Control
CI/CD
settings file