Claude Fable 5.2 Data Retention, DPDP and Claude Code Security

Does Claude Fable 5.2 support zero data retention?

By default, no. Anthropic's retention documentation says Claude Fable 5.2, Claude Mythos 5.2 and the earlier 5.1 and 5 versions "require 30-day data retention and are not available under ZDR unless expressly authorized by Anthropic." Anthropic calls these Covered Models. Its help article adds that prompts and completions are kept for at least 30 days and then deleted automatically, and that automated safety systems assess what is kept. A request from an organisation without the required setting returns a 400 invalid_request_error asking it to enable data retention. Anthropic gives safety monitoring as the reason.

• •Longer holds apply under a safety investigation or a legal requirement.
• •Flagged content is a separate rule that covers every model: even under ZDR or HIPAA, Anthropic may retain inputs and outputs for up to 2 years if a chat or session is flagged.
• •ZDR for Covered Models is unavailable in workspaces, Claude Enterprise organisations and third-party platforms, according to the help article.
• •The exception: eligible customers can use the Fable models under ZDR for internal business applications, for a limited time, as a transition to Enterprise Frontier Safeguards. Anthropic may modify or withdraw that arrangement.
• •Enterprise Frontier Safeguards runs automated monitoring on data stored in your own cloud account. On 9 November it was in private beta, with Bridgewater, Salesforce and Stripe named. We found no general availability date.

How should you structure workspaces around the 30-day rule?

Sort work by data class first and pick the model second. Anthropic's own guidance is to start most work on Claude Opus 5.5, which is outside the 30-day rule, so routing sensitive data away from Fable costs you little on most tasks. Fable 5.2 can then take the long engineering jobs where nothing personal is in the prompt.

Can Claude inference stay in India?

On Anthropic's own API, no. The inference_geo parameter accepts only global and us, and workspace geo, which controls storage at rest, offers only us. India appears only through the cloud platforms. Anthropic's regional compliance page lists Asia-Pacific residency on Bedrock and Vertex, names India in that group, and names Sonnet 5, Opus 5 and Haiku 4.5 in its FAQ. AWS published an India profile on 29 September for those three models, routing between Mumbai and Hyderabad, and says it "keeps inference within India." Fable 5.2 is different. AWS's model card, read on 11 November, shows no In-Region or Geo option for Mumbai or Hyderabad, only Global, and the sole Geo profile is US. It also requires setting the data retention mode to aws_review. We could not confirm Fable's India options on Google Cloud. If a contract or regulator requires inference in India, Fable 5.2 on Bedrock does not meet it today.

• •Worked cost of US-only inference on Fable 5.2 at $8 input and $40 output per million tokens, for 10 million input and 2 million output tokens: (10 x $8) + (2 x $40) = $160, about Rs 15,017 at the RBI reference rate of Rs 93.8579 per USD on 9 November 2026.
• •With inference_geo set to us the multiplier is 1.1x: $176, about Rs 16,519, so Rs 1,502 extra for this volume.
• •The same volume on Opus 5.5 at $4 and $20 costs $80, about Rs 7,509, before any difference in how many tokens each model uses.
• •Enforce the choice in workspace settings with allowed_inference_geos and default_inference_geo, and log the inference_geo field returned in the usage object as evidence.

What does the DPDP Act change for Claude users, and when?

Very little this week and a great deal in six months. Trackers differ by a day on the exact dates, so we give them as mid-November 2026 and mid-May 2027. The mid-November phase brings in only the Consent Manager provisions, so on our reading a business that is not applying to be a Consent Manager gets no new duty from that phase. The mid-May 2027 phase brings in the rest, including notice, consent, data fiduciary obligations, breach notification and the section on transfers outside India. This section is general information and not legal advice, so take your own contracts and notices to counsel.

• •You stay liable. The Future of Privacy Forum's analysis says the Act places obligations on data fiduciaries, who may engage processors only under a valid contract. Anthropic's docs say it is the processor on the Claude API, Claude Platform on AWS, Anthropic-managed Google Cloud and Foundry, while the cloud provider is the processor on Bedrock and Google-managed Google Cloud. Your contract file should match your route.
• •Your deletion promises must allow for vendor retention. If you send personal data to Fable 5.2, a consent withdrawal cannot produce deletion at the vendor in under 30 days, and a flagged session may be held for up to 2 years.
• •On the Future of Privacy Forum's 2023 reading, cross-border transfer works as a blacklist. Transfers are allowed unless the government restricts a named country or territory, so on that reading the Act alone does not bar global inference. AZB lists cross-border transfer requirements in the 2025 Rules too, so have counsel check those and any notifications before May 2027, and check your sector regulator separately.
• •Security safeguards and breach notification begin in that phase too, so start logging which workspace and model handled personal data.

What do Claude Enterprise inference hooks do, and what do they miss?

Inference hooks send each governed prompt, and optionally tool calls, to an HTTPS security server that you or your vendor runs, and wait for an allow or deny. Anthropic's documentation says "a denied request never reaches the model." They are generally available on Claude Enterprise as of the 5 November announcement, and one hook covers claude.ai, Cowork, Claude Code, Claude for Microsoft 365, Claude Tag and Claude in Chrome. Only the Owner and Primary owner roles can configure them. DLP is the most common use.

• •Claude Platform organisations are out of scope, and hooks are not available on Amazon Bedrock or Google Cloud. API products need your own gateway check before the request leaves your network.
• •Verdicts are allow or deny only. Rewriting or redacting a prompt is not supported.
• •Raw file and image bytes are never sent, so a screenshot of a document is not inspected.
• •Anthropic warns not to rely on hooks as your only control for voice.
• •A denied message is resent with every later message, so users must edit it out or use /rewind in Claude Code. Train people on this before you enforce.

What else changed for Enterprise admins in November?

Three smaller releases from 5 and 6 November affect how you evidence control and charge back cost.

• •Security center (beta) shows each security setting against Anthropic's recommended value and produces auditor reports. Anthropic says it is "not a certification or a guarantee of compliance with any law or standard."
• •Project tags (beta, 6 November) attribute spend to a charge code or client-matter number, with up to 16 tags per project. They are unavailable to organisations using ZDR, customer-managed keys or data stored outside the US, and do not cover the Claude API. A ZDR organisation needs another way to bill clients.
• •For Claude Code, tags need attributionProject in .claude/settings.json and version 2.1.305 or later.
• •Member sign-out through the Admin API (beta) does not end API keys, cloud sessions already running or scheduled routines. Your offboarding checklist must revoke those separately.

How did 1,122 Claude Code settings files end up leaking credentials?

According to Dataconomy's report of research by Apiiro, the "Yes, and don't ask again" option can save a full approved shell command, including any embedded password or key, into .claude/settings.local.json. Apiiro found 1,122 publicly accessible settings files containing credentials, after examining 104,185 public copies across 103,617 GitHub repositories. About 1.1% held what it classed as real secrets, or 1,379 distinct credentials. Apiiro did not test them. Anthropic's Boris Cherny said on 7 November that built-in credential scans were being considered, with no date given. Claude Code excludes the local file from git when it writes that file itself, but a hand-created file gets no such protection and the shared file is meant to be committed. Run these checks on every repository and rotate anything you find.

What should a Claude Code managed settings file contain?

Managed settings sit above every user, project and local file, so company policy belongs here. allowManagedPermissionRulesOnly matters most for the Apiiro problem: it hides the always-allow choices in permission prompts and stops new rules being saved, so your allow list must live here. Place it at /Library/Application Support/ClaudeCode/managed-settings.json on macOS, /etc/claude-code/managed-settings.json on Linux and C:\Program Files\ClaudeCode\managed-settings.json on Windows, then run /status to confirm. A file on the device does not reach cloud sessions, which need server-managed settings from the admin console. The deniedModels block keeps Fable off developer machines until retention is settled, and needs version 2.1.283 or later.

Which governance mistakes are easiest to make?

How Tech Arion helps

Tech Arion builds with Claude Code and advises Indian teams on model routing, workspace design and agent policy. We can map your data classes to routes and draft your managed settings. Your lawyers keep the final word on DPDP. See /services/ai-consulting, or read /blog/claude-fable-5-2-system-card-plain-english-findings for what the system card means for agent deployments.

Frequently asked questions about Claude Fable data retention and security