🧭 Claude Code v2.1.271: Per-Command Domain Sandboxing, Fast Mode in Remote Sessions, and a Critical 64 MiB Config Fix

Released September 14, Claude Code v2.1.271 ships four changes that matter most for teams running auto mode on Remote sessions and self-hosted runners. The headline addition is per-command allowed_domains sandboxing for Bash, PowerShell, and Monitor tool calls: in auto mode with sandboxing enabled, each command's outbound network access is now restricted to a per-call domain allowlist rather than the runner-level blanket list. A misconfigured tool can no longer reach domains it was never supposed to touch, even if the runner's global configuration would have permitted it.

What changed in v2.1.271

• Per-command allowed_domains: Set a domain allowlist on individual Bash/PowerShell/Monitor calls in auto mode. Traffic outside the allowlist is blocked at the command boundary — not just logged.
• Fast mode in Remote sessions: The host fast_mode setting or the /fast toggle now propagates to Remote cloud and self-hosted runner sessions, provided the organisation policy allows it. Previously, fast mode was silently ignored for any session running outside the local machine.
• --drain-marker-file for self-hosted runners: Passing a file path via this flag tells the runner to write a marker file when it begins draining active sessions before shutdown. CI orchestration scripts can poll for this file instead of relying on timing heuristics to know when the runner is safe to terminate.
• Critical bug fix — 64 MiB host config directory: If the host configuration directory (settings, skills, plugins, MCP servers) exceeded 64 MiB, all of it was silently discarded. Sessions appeared to start normally but ran without any of the organisation's configuration. The fix raises the limit and adds a warning log entry if the directory is large.

🧭 Anthropic Launches Claude for Financial Advisors — Pre-Built Wealth-Management Skills Across Eight Platform Integrations

Anthropic has launched Claude for Financial Advisors, a vertical product that connects Claude to the wealth-management software advisors already use daily. Rather than asking firms to build their own Claude integrations from scratch, the product ships pre-configured skills mapped to common advisor workflows: client research, meeting preparation, portfolio review summaries, documentation drafts, and compliance pre-screening. The launch covers eight platforms on day one — BlackRock Aladdin, Charles Schwab Advisor Center, Addepar, Envestnet, iCapital, Orion, Wealthbox, and Redtail Technology.

How it fits into Claude Co-Work

Claude for Financial Advisors is an add-on to an existing Claude Co-Work plan, priced at $70–$120 per user per month depending on the platforms enabled. Each integration installs as a connector within Co-Work, so advisors can access it through the same chat and canvas interface they use for everything else — no separate application or browser tab. Investment recommendations remain subject to mandatory human review and cannot be transmitted to clients without an advisor sign-off step built into the workflow.

Compliance guardrails baked in

• Fiduciary framing: Every portfolio summary includes a disclaimer that Claude's output is analysis support, not investment advice, and is not to be used as the sole basis for client recommendations.
• Compliance pre-screening: Before a document leaves Co-Work for client delivery, a compliance skill checks it against the firm's flagged-term list and regulatory language library. Violations surface as inline suggestions, not blocking errors — the advisor decides.
• Audit trail: Every Claude interaction that touches client data logs to the firm's existing record-keeping system via webhook — satisfying SEC books-and-records requirements without a separate archiving integration.
• Zero training on firm data: Standard zero-data-retention terms apply; client portfolios and meeting notes are never used to train Anthropic models.

🧭 Project Glasswing Expands: EU's ENISA Becomes First European Agency with Sanctioned Access to Claude Mythos 5

Anthropic has expanded Project Glasswing — its structured access programme for government cybersecurity agencies — to include the European Union Agency for Cybersecurity (ENISA). ENISA becomes the first EU-level body to receive sanctioned access to Claude Mythos 5, the most capable and most restricted model in Anthropic's lineup. Britain's NCSC had been the sole non-US Glasswing participant since the programme launched alongside Mythos in May 2026; ENISA's inclusion means the EU can now use Mythos to support its NIS2 directive work, critical infrastructure vulnerability research, and AI incident response without routing requests through a third country.

Why Mythos access is restricted in the first place

Claude Mythos 5 is not generally available. Anthropic gated it from launch because evaluations showed it can identify novel software vulnerabilities in critical infrastructure — water treatment SCADA systems, power grid controllers, financial settlement networks — at a level that would give a sophisticated attacker a meaningful advantage if the model were misused. Glasswing provides access under a legal framework that requires participating agencies to use Mythos only for defensive purposes, prohibits re-distribution of outputs to commercial vendors, and gives Anthropic's safety team audit rights over usage logs. ENISA agreed to the same terms as the US Cybersecurity and Infrastructure Security Agency (CISA) and NCSC.

What ENISA can now do with it

• Vulnerability triage at scale: Mythos can ingest national-level CVE disclosures and triage which affect NIS2-regulated entities within the EU, a task that previously required weeks of manual analysis per quarter.
• Threat actor attribution support: Correlating malware code signatures, infrastructure reuse patterns, and TTPs across ENISA's member-state CERT feeds — Mythos handles multilingual source material across all 27 EU languages natively.
• AI incident response simulations: Mythos can model adversarial AI use-cases that ENISA is tasked with preparing defences against, including synthetic disinformation campaigns and automated spear-phishing pipelines.