🧭 Enterprise Frontier Safeguards: Zero Data Retention Now Comes with Misuse Detection in Your Own Cloud

Anthropic has shipped Enterprise Frontier Safeguards (EFS), resolving a genuine tension that has blocked regulated organisations from moving to Claude's most capable models. The problem: security teams need zero data retention (ZDR) — no prompt or agent transcript sitting on a vendor's server — but safety teams need misuse detection, which historically required the vendor to hold those same logs long enough to correlate suspicious patterns. EFS breaks the deadlock by routing activity logs to cloud infrastructure controlled by the customer (AWS S3, Azure Blob Storage, or GCS under the customer's own encryption keys), while Anthropic operates detection logic against that data without any human ever reading the content.

How EFS works in practice

• Customer-held storage: Every prompt, completion, and agent action log is written directly to a bucket or blob container in the customer's cloud account. Anthropic never stores a copy.
• Anthropic-operated detection: Anthropic's misuse-detection models run against the customer-held data using a cross-account access pattern — the same architectural approach used in security-information and event management (SIEM) systems. No human analyst at Anthropic reviews the content; only automated signals and aggregated risk scores are surfaced.
• Encryption ownership: The customer controls the KMS keys. If the customer revokes access, detection stops. This gives legal and compliance teams a clean audit trail showing that Anthropic cannot read their data unilaterally.
• Fable 5 + 5.1 ZDR now: While EFS phases in through autumn 2026, eligible enterprise customers immediately receive zero data retention on Fable 5 and Fable 5.1 as a bridge guarantee.

Who developed this and why

Anthropic worked with more than 100 customers in financial services, healthcare, manufacturing, telecom, law, retail, and the public sector to validate the architecture. The three major cloud providers (AWS, Google Cloud, Microsoft Azure) were involved to ensure the cross-account pattern works reliably at production scale across each hyperscaler's permissions model.

🧭 Claude Code v2.1.261: /skill-doctor, Subagent System Prompt Files, and Organisation Policy in /status

Claude Code version 2.1.261, released September 4, is a focused patch with several features that will save time for anyone managing Claude Code at team or organisation scale.

New commands and flags

• /skill-doctor — Lists all skills that are currently loaded into your Claude Code session but have not been invoked during that session. Run it when you suspect a skill is misconfigured or shadowed by another skill. The output shows the skill name, file path, and trigger pattern, making it straightforward to diagnose why a skill that should be firing isn't.
• --append-subagent-system-prompt-file <path> — Reads an additional system prompt from a file on disk and appends it to the subagent's system prompt at launch. Useful when you need to inject organisation-specific instructions or project context into subagents without embedding them in the parent agent's prompt or in a hardcoded flag string. The flag accepts an absolute or repo-relative path; the file is read once at subagent startup.
• Organisation policy in /status and claude doctor — Both commands now display which organisation-level policy settings are active for the current session. If your administrator has set managed values for things like managedMcpServers, disableNonEssentialTraffic, or model restrictions, you'll see them listed alongside your personal settings. This eliminates the "why is this setting not taking effect?" debugging cycle when personal and org-level settings conflict.
• bashOutputMaxChars and taskOutputMaxChars — Two new settings that raise the character-limit ceiling for bash command output and task output respectively. Previously, very long stdout from shell commands or task completions would be silently truncated. Setting these in your ~/.claude/settings.json (or org-level settings.json) lets you capture the full output for log-heavy scripts or build pipelines.

How to use bashOutputMaxChars and taskOutputMaxChars

// ~/.claude/settings.json — raise output limits for log-heavy workflows
{
  "bashOutputMaxChars": 200000,
  "taskOutputMaxChars": 100000
}

🧭 Anthropic Opens $5 Million Grant Programme for Independent AI Wellbeing Evaluations

Anthropic has announced a $5 million grant programme to fund independent researchers building open-source evaluations that measure how AI systems affect user wellbeing. The programme is explicitly designed to address a gap that Anthropic's own Safeguards team has flagged internally: for contexts such as AI companionship and mental-health support, no widely accepted, rigorous measurement standard currently exists. Without validated external evaluations, both developers and regulators are making decisions based on anecdote or internal benchmarks with obvious conflicts of interest.

What grantees will build

• Open-source evaluation suites — all evaluations, datasets, and related tooling must be released openly, so any developer or AI provider can run them
• Focus domains: companionship and mental health — these are the two areas where Anthropic's Safeguards team says the risk of harm is highest and the measurement gap is widest
• Multi-turn, realistic conversations — Anthropic's published guidance for grant applicants specifies that evaluations should reflect realistic escalating conversations, not single-turn prompts, because risk in these contexts typically develops over a session
• Expert validation — evaluations must involve clinical or domain experts in their design and be validated against expert judgement, not just automated metrics
• Both overcompliance and overrefusal tested — a rigorous wellbeing evaluation should catch models that are too accommodating (enabling harm) and models that refuse legitimate emotional support (a different kind of harm)

Key dates

• Application deadline: September 21, 2026
• Notification for full proposals: October 5, 2026