Archive — 2026-10Week 3

Anthropic announced physicist Christopher Stubbs has joined the Long-Term Benefit Trust (LTBT) — posted 2026-10-21 in Anthropic News. Stubbs is a former Dean of Science at Harvard (2018 to 2024) and the Samuel C. Moncher Professor of Physics and of Astronomy. He was on the High-Z Supernova Search Team that discovered the accelerating expansion of the universe in 1998, served as inaugural project scientist for what became the Vera C. Rubin Observatory, and is a JASON member who has advised the US government on national security and arms control. He is the Trust's first physicist. The LTBT can appoint a majority of Anthropic's board; trustees are independent of management and investors and advise the board and leadership on high-stakes decisions involving AI risks and societal impacts. Other trustees include Neil Buddy Shah (Chair), Richard Fontaine and Ben Bernanke. The announcement came the same week as Anthropic's first Public Benefit Report, and, subject to stockholder approval, the company has committed to setting aside a portion of its shares for the Anthropic Public Benefit Fund.

Claude Code v2.1.306 — Fixed: `claude -p` inside cloud sessions reporting "Not logged in" when started without the session's environment variables; an API error being replaced by "prompt is too long", with an unneeded compaction, in sessions with many MCP tools.

Anthropic published "Using Claude to reduce the attack surface of code" — posted 2026-10-20 on Anthropic Research by Chris Rohlf of the Frontier Red Team. Claude analyzed 25 open-source projects (C, C++, Java, PHP, Python; over five million lines), inventorying more than 17,000 configuration options (3,842 security-relevant) to find code needlessly reachable in default configurations, option combinations that cancel security controls, and code running before authentication, 493 issues in total. All 11 gated projects had unintentionally reachable code: 93 unnecessary pre-authentication surfaces in defaults, 37 places where code ran or state was recorded before authentication, and 61 option combinations that silently weakened an administrator's pre-auth control. 162 sensitive combinations were accepted with no error or documentation, 94 of which silently disabled a security control; one client-certificate check conflict had existed for 21 years. OpenSSH case: about 4,500 lines of experimental ML-DSA-44 post-quantum signature code were reachable by unauthenticated clients in the default sshd, roughly 16% of pre-auth code; 10.5 moved the check earlier and 10.6 placed allow-list checks inside the parser, removing about a fifth of the previously reachable code. Anthropic packaged the technique as a reusable Claude skill and is working to integrate it into OSS-Scanner.

investingLive: Anthropic IPO draft shows nearly 15-fold revenue jump and over $600 billion in compute commitments — Eamonn Sheridan, 2026-10-20, citing a draft prospectus reviewed by the Wall Street Journal (Anthropic declined to comment; not confirmed by the company). H1 2026 revenue was about $16.3 billion (about $1.1 billion a year earlier; about $4.6 billion for full-year 2025). Net loss was about $81 billion (about $13 billion a year earlier), of which roughly $72 billion came from accounting charges tied to convertible notes and other financing instruments; full-year 2025 net loss was about $42 billion. Compute and infrastructure spending was nearly $9 billion (about $2.3 billion a year earlier). Infrastructure commitments exceed $600 billion, including about $450 billion in cloud agreements over seven to 10 years, mostly non-cancellable (Google about $111 billion, AWS about $110 billion, Microsoft about $31 billion), up to $86 billion of Nvidia chip capacity via Nscale and Lambda, and over $160 billion to lease Google chips through Broadcom; deals under negotiation could add up to $150 billion. Two customers accounted for about a fifth of H1 revenue. The article gives no specific IPO date and says the listing Anthropic is preparing could value it at up to $2 trillion.

Claude Code v2.1.305 — cumulative update. Added: `sandbox.allowNestedUserNamespaces` lets sandboxed Linux commands create user namespaces (Chromium, rootless podman), which lets them get around the sandbox's file protections; `CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS` lets unattended retry mode (`CLAUDE_CODE_RETRY_WATCHDOG`) notice a usage limit lifted before its reset. Fixed: every request failing on Vertex, Bedrock and other providers when the endpoint or a gateway rejects tool search (a 2.1.304 regression), now retried without it; `/effort` and `/autocompact` not applying when the model name differed from a `modelOverrides` value only by letter case or a `[1m]` tag; new worktrees in a bare repository sometimes starting from local HEAD instead of the default branch, and unchanged subagent worktrees not being cleaned up; Bash on a full disk failing with a raw `ENOSPC` path, now saying the disk is full and nothing ran; an MCP server passed by a cloud session or the SDK's `setMcpServers` later replaced by a same-named server from user settings; Claude naming the wrong model after a subagent switched to a fallback model; cloud sessions and teammates not being told when compaction left out the oldest messages; scheduled Cowork cloud sessions waiting five minutes per subagent whose WebFetch needed an unattended URL approval; Workflow resume re-running completed `pipeline()` and `parallel()` agents instead of serving them from the run journal; Remote Control leaving empty "(session)" entries in `/resume`; `/usage` rows cut off in narrow terminals; startup warnings printing as raw text above the logo, now sent to the debug log; screen reader mode search labels and highlighted-text spacing; plus several Self-hosted runner fixes (`Resource temporarily unavailable` when `TERM` is set, repository access failing since 2.1.296 when credentials come through `GIT_CONFIG`, spawn-runner hook falsely timing out). Improved: `/usage` warnings are now one-line fixes naming their source, applied with `f` (also in VS Code and `/doctor`); MCP consent dialogs are less likely to treat a stray Enter as Accept; a mod's `Image` draws in iTerm2 3.5+ fullscreen rendering; Remote Control spawn-mode messages. Changed: `allowedMcpServers` no longer filters `managed-mcp.json` stdio servers that use `${VAR}` only in `env`, so ones your allowlist left out now load on upgrade.

Anthropic published "Claude scorpion toxin binder design" on its Research page on 2026-10-20. The life sciences team worked with investment research firm Citrini Research and asked an internal research model (running in Claude Science) to design de novo binders against scorpion neurotoxins, using the same prompt as the August protein minibinder study (modified with Citrini). Claude worked largely autonomously over 64 hours (24 hours after the first prompt, with four follow-ups), using 12 design tools to generate 318,484 designs and co-fold and screen 553,052 design-toxin pairs, selecting 180; the team added 31 more with Claude's help, for 211 in total, using about 10,000 GPU hours. Targets were five sodium-channel neurotoxins from Mexico (Cn2), Brazil (Ts1), North Africa (AaH I, AaH II) and the Middle East (LqhIII), with single-toxin, β-family, α-family and all-five binder goals. Adaptyv Bio (cell-free expression) and Twist Bioscience (mammalian cells, Fc fusions) measured binding by surface plasmon resonance: 17 of 211 designs met the binder criteria, all against a single toxin (7 Cn2, 4 AaH II, 4 AaH I, 2 LqhIII); 11 of 60 single-toxin designs were binders (18%), consistent with typical campaigns, and no multitoxin design met the criteria for more than one toxin. Anthropic has not confirmed neutralization of venom in a cellular environment and does not claim an antivenom. Outside experts Susana Vázquez Torres and Timothy P. Jenkins called it a promising first step; all 211 designs and data are released on Hugging Face with a technical report.

Claude Code v2.1.304 — cumulative update. Added: `/auto-mode-setup`, a guided setup that reads the project and recent sessions to teach auto mode about the environment (also on Bedrock, Vertex, Foundry, Claude apps gateway and telemetry-disabled sessions); `/claude-api managed-agents-secure`, which compares a Managed Agents agent's settings with Anthropic's published security guide read-only; a warning on Pro and Max plans when resuming an inactive conversation would use a noticeable share of the 5-hour limit; `decider: "person"` and `e.isReadOnly` for a mod's `tool.check`; and prior tool names in the compaction summary of the transcript view (Ctrl+O). Fixed: `claude --worktree` removing the worktree and branch without asking at exit when HEAD was detached or on another branch, which could lose commits; `rm -r` of a mounted drive or a top-level `/Volumes`, `/media`, `/mnt/c` folder skipping the safety prompt; `CLAUDE_CODE_TRANSCRIPT_LOCAL_GC` dropping messages when the disk filled; `CLAUDE_CODE_AUTO_COMPACT_WINDOW=200k` silently setting a 100K window (now accepts `k` and `M` suffixes); `context_management: Extra inputs are not permitted` errors behind LLM gateways; a plugin keeping a headless run going past `--max-turns` by spawning subagents; Remote Control failing to start from Claude Desktop when DISABLE_TELEMETRY or DO_NOT_TRACK is set; plus several VSCode and Cowork cloud session fixes. Changed: tool search is on by default behind a custom `ANTHROPIC_BASE_URL` (retried without it on a 400 or 422); the skill list is resent after compaction; one skill source can register at most 1,024 hooks. Code Review: added Quick, Standard (default) and Deep review levels, configurable per organization or per repository, and an option to review every repository in a GitHub organization including ones added later.

Claude Code v2.1.303 — the official release note lists only bug fixes and reliability improvements, with no specific items itemized.

Anthropic published "Investing in the public benefit" on 2026-10-18 in the official Newsroom. Subject to stockholder approval, the company will commit 5% of its shares to a newly established public benefit fund; the board may add 2% for every 50% increase in company value above a baseline, with neither the baseline nor the valuation milestones disclosed — Anthropic says it will share details on the fund's governance and structure in the coming months. The fund is structured as a wholly-owned LLC (not a 501(c)(3)), preserving flexibility between forms of giving; the equity it holds will not be used for lobbying or political giving, in the US or abroad. Use cases include large donations to traditional nonprofits and direct distributions to specific populations, with labor market impacts cited as the first example — Anthropic says bolder approaches may be needed beyond its existing $200 million commitment to economic futures research. Work kept inside the company: AI safety and interpretability research, applied biomedical research, and discounted Claude access for nonprofits and scientific institutions, since these carry both public benefit and for-profit value. Context: co-founders pledged to donate 80% of their personal equity at founding, and an equity donation match program covers about half of employees; combined with the new 5% commitment, roughly 20% of Anthropic shares will be dedicated to the public good, with valuation-milestone additions counted separately. The commitment redirects equity that would otherwise go to executives as escalating compensation awards. The post does not mention an IPO; Anthropic also plans to publish an annual public benefit report on progress against its mission.

Claude Code v2.1.302 — cumulative update. Added: `CLAUDE_CODE_AUTO_MODE_MODEL` to choose the model used by the client-side auto mode classifier (server-side auto mode unaffected); an Idle compaction row in `/config`, with post-compaction notices now explaining how to turn it off; a notice in `/design-sync` and `/design-login` that standalone Claude Design (claude.ai/design) closes on 2026-12-14. Improved: faster startup with many plugins enabled; the server-side auto mode classifier now compacts a conversation too long for the safety check instead of denying every checked tool call. Changed: `claude -p` and Agent SDK sessions with no permission mode start in default mode on third-party providers, with telemetry off, or on a fresh install (`--permission-mode auto` keeps auto); fork subagents (`subagent_type: "fork"`) are on by default in `claude -p`, the Agent SDK, the desktop app and VS Code (`CLAUDE_CODE_FORK_SUBAGENT=0` turns them off); the desktop app and VS Code always run the main conversation's subagents in the background; at most 12 prompt or agent hooks of one event run at a time. Fixed: "Prompt is too long" when stopping a turn during compaction; requests failing after switching to Vertex or Bedrock mid-session; the Read tool loading image files of any size into memory (sources over 128 MB are now refused); prompt cache misses when a tool added mid-session was reloaded; plugin subcommands exiting before a queued settings write finished; plugin marketplaces re-cloning on every refresh under git's reftable format; worktree isolation refusing worktrees under git 2.56 or later; MCP servers showing no tools when negotiating protocol version 2026-07-28 but omitting `resultType`, `ttlMs` or `cacheScope`; cloud session cost and usage restarting from zero after resume; every file write refused on exFAT or FAT32 drives on Windows; plus many fixes for OpenTelemetry, the PowerShell tool, VS Code, Remote Control, Claude Tag (Slack), Code Review setup and cloud routine status.

Anthropic publishes "The AI expectations gap" — posted 2026-10-16 on Anthropic Research. Compares US adults (Morning Consult, 8/11–8/23, 10,980 weighted respondents) with Anthropic staff (internal survey 8/9–8/12, 1,334 respondents) on expectations for AI. Nobel-level scientific discovery by 2030: 39% of US adults vs. 91% of staff. Share of automatable work done with AI by 2030: 65% (staff) vs. 42% (US adults). Share of own work AI will do by 2030: median 75% for staff vs. 28% for US workers; over half of staff expect 80% or more vs. about 5% of US workers. Unemployment forecasts: US workers expecting AI to do 80%+ of their work forecast about 12.2% unemployment in 2030 (above the 2008 financial crisis peak) versus the current 4.2%; each 10-point rise in work exposure corresponds to about 0.4 points more forecast unemployment. Staff median estimate of AI's share of frontier-model work: 51% today, 80% by end-2027, 95% by end-2030.

Anthropic publishes "How employment is changing at Anthropic" — posted 2026-10-16 on Anthropic Research. Discloses the job-structure shift under widespread AI agents: junior-level share in technical departments fell from 21% to 14%, the junior technical hire share fell from 27% to 9%, and technical departments' overall share fell from 61% to 56%. In March 2026 agent work hours exceeded human hours for the first time, reaching six times by August. Engineer code commits are about 5x their level six months earlier, with no higher deletion or revert rate; roughly 80% of interactive Claude session time is idle waiting for human input; meta-work managing agents and tooling grew from 10% (fall 2025) to 22% (spring 2026) of interactive sessions. Interview comments on AI tool use roughly tripled for technical roles and quadrupled for non-technical roles vs. 2024. Anthropic estimates revenue per employee is 2–4x Meta's and Google's at the same stage, and recommends stock ownership to share AI's gains, an early warning system for internal reallocation, and notes government retraining and income support are still needed.

Anthropic publishes "Designing more efficient quantum circuits with Claude" — posted 2026-10-16 on Anthropic Research. The team gave Claude plain-language specifications for 8 quantum circuits (chemistry, optimization, quantum simulation, number theory, numerical computing) with an automated evaluator checking correctness and scoring. Working largely autonomously, Claude produced designs using fewer qubits and fewer gates than the best previously published versions, with some improvements of an order of magnitude or more. Example: elliptic-curve point addition took 53 hours and yielded a reversible divsteps that stores no branch history; one design uses 12% more qubits but 17x fewer gates. The DQI (quantum optimization decoder) Reed-Solomon decoder uses about 1/12 the gates of Google's previous construction with 19 qubits of scratch (vs. a worst-case reserve of 665 qubits); testing 403M sample decodings showed about one run in 10,000 would exceed that reserve. FeMoco (nitrogenase cluster) context: Toffoli gate counts dropped from about 50 trillion (2017) to about 340 million, roughly a 150,000x reduction (over 8 years and multiple research groups).

Anthropic Frontier Red Team publishes "Claude reveals ECDSA curve origins" — subtitled "It's just timestamps," published on 2026-10-16 in Anthropic Research. Authors: John Wood, Henry de Valence, Ava Howell, Zack Voase, Lucy Sweet, Amir Omidi, Paul Crowley, Iain McGinniss, Nicholas Carlini, and Steve Weis. Context: the seed values for NIST elliptic curves P-192, P-224, P-256, P-384, and P-521, published in FIPS 186-2 in 2000 without explanation, had long been presumed to come from the NSA. Finding: the seeds are not hash outputs but outputs of the C standard library's `lrand48` generator, seeded with a Unix timestamp. Claude identified the generator by spotting statistical patterns such as all-zero top bits, and recovered internal states using the LLL lattice reduction algorithm. The prime-curve seeds correspond to timestamps between May 1997 and April 1998; the binary-curve seeds use half the timestamp, mapping to 1998-04-17. P-256 used candidate 1,842 of its generator run; candidate 876 of the same run also yields a valid curve, but the published seed skipped it. Claude also traced the base points G of P-192 and five ANSI X9.62 example curves to Arjen Lenstra's FreeLIP package random number generator seeded with small integers such as 20000. Time: Claude located the P-192 through P-521 seed origins in under 30 minutes; several Anthropic staff had previously chased a rumored humorous passphrase lead without success, and the article says the lead was wrong. Unresolved: the seed for binary curve B-163; the base points G for P-224, P-256, P-384, P-521, and the NIST binary curves; and why some valid candidates were skipped. Bounty: $36,864 (set by cryptographer Filippo Valsorda), to be donated to the Internet Security Research Group; the reproduction code is released under CC0. Security implications: the authors emphasize this is "not a cryptographic break, new attack, or weakness" in ECDSA; systems using P-256 and other NIST curves need no changes, and the result makes backdoor theories less plausible — the seeds came from a simple and publicly reproducible generator.

claude.com publishes "How an Anthropic Japan BDR runs his day with a fleet of Claude agents" — by Shahboz Ibragimov, a business development representative on Anthropic's Japan team, posted in Resources on 2026-10-16 (about a 14-minute read). The opener describes mornings previously spent on manual lead routing, email drafting, and CRM updates, now mostly handled by a fleet of Claude agents before the workday starts. The prototypes were built inside Claude and then moved onto Claude Managed Agents (beta) so they run on schedule, share a memory store, and use credentials in managed vaults. The fleet is five single-purpose working agents, each with a Japanese name, plus a coordinator agent that updates the others' instructions when he asks for changes in Slack: Furiwake sorts the day's inbound leads to reps; Sabaki qualifies his assigned leads, updates Salesforce, and drafts first-touch emails; Henji drafts replies to customer emails hourly in Japanese or English; Shoudan keeps Salesforce opportunity records current; Nokori summarizes meetings at day's end and drafts follow-ups. Tasks needing further work (escalations, procurement questions) are handed off to Claude Tag in Slack, which can set reminders. Models are matched to task complexity: Furiwake and Shoudan run on Sonnet, Sabaki, Nokori and the coordinator run on Opus, and Henji runs on Sonnet with Opus as an advisor for hard calls. He estimates the fleet saves about 60 hours a week. Lessons: describe each job once; invest in a shared knowledge base of corrections; hand agents more responsibility sooner than feels comfortable.

Claude Code v2.1.301 — cumulative update. Added: `CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS`, so unattended retry mode (`CLAUDE_CODE_RETRY_WATCHDOG`) ends a request at a usage limit that resets too far away instead of waiting; OpenTelemetry tracing guidance in the `claude-api` skill (including setting up trace destinations with the `ant` CLI); the `%` motion in vim mode (jump to matching bracket, alone or with d/c/y); `rows` on a mod's `Input` (multi-line text field; Enter submits, Shift+Enter or backslash+Enter inserts a newline); `classifier_verdict`, `permission_mode` and `classifier_category` (with `OTEL_LOG_TOOL_DETAILS=1`) on OpenTelemetry `tool_decision` events to tell auto mode classifier decisions from permission rules; VS Code LaTeX math rendering (`$$`, `\[ … \]`, inline `$…$` and `\(…\)`) and a "Get more usage" link on the usage-limit banner for Pro and Max. Changed: switching to a model with a smaller context window now has the previous model summarize an over-long conversation first; a mod hook's `.catch` runs only when the hook throws or times out; a mod's `$.env.set` resolves with `{}` on success or `{ deny }` with the reason; subagent progress trail now updates incrementally for better rendering performance. Fixed: the `fable` (and `best`) alias erroring when the newest Fable model is unavailable to the organization (now retries once on the previous Fable model); quitting removing a `WorktreeCreate` hook worktree without asking when it may hold unpushed commits; a misleading "Please run /login" error on a full disk; Grep and Glob reporting no matches when the search program failed; sign-in refresh timeouts and sign-outs after short commands such as `claude auth status` exited mid-refresh; conversations compacted repeatedly behind an LLM gateway that names a different model; the Read tool reading binary files such as .xlsm, .parquet, .epub and .heic as text; skills not listed to the model after `/rewind`; vim mode staying in INSERT after Escape; plus various terminal rendering, self-hosted runner, Windows and VS Code extension issues.

Anthropic published "An update on independent evaluators" on 2026-10-15 in the official Newsroom. A six-person METR (Model Evaluation and Threat Research) team now has employee-like access at Anthropic: company hardware, office space, and access to internal tools and systems roughly on par with Anthropic's own risk assessment teams; employees are encouraged to share information, with only three narrow carve-outs — customer data, attorney-client privileged material, and model architecture details. Anthropic cannot veto METR publications but can require redactions in those categories, and METR may publicly state whether it considers the redactions justified; other METR staff are walled off from the team. METR will publish analyses of Anthropic's risk levels, incidents, and practices; the two have co-reviewed several Anthropic Risk Reports over the past year. SecureBio is currently reviewing the chemical and biological weapons chapter of the August 2026 Risk Report, having previously published a pilot review of an earlier report. Accenture and Faculty are moving toward a broader partnership covering model evaluation, red-teaming, alignment assessments, safeguard testing, and capability building. Hugging Face is in early discussions. Anthropic lays out three evaluator-access tiers: (1) narrowly scoped investigations with Anthropic approval required before publication; (2) broader information access with publication without Anthropic approval, subject to the same redaction rules; (3) employee-like access (METR's current arrangement). Background: Anthropic committed last month to stronger independent third-party assessment of its models and safety practices, and signed the White House Accord on Super Intelligence pledging to work with independent external auditors or evaluators; the article draws an analogy to bank supervision, where regulators place examiners inside large banks. Anthropic quotes: "We aim to close the gap between what evaluators understand" and "We're embedding evaluators voluntarily, and we believe every frontier AI company should do the same." No funding amounts or contract durations are disclosed.

claude.com publishes "Code Review is now generally available" — posted in Resources on 2026-10-15. Claude Code's Code Review leaves research preview and is available to Claude Team and Enterprise organizations. Reviews now run about three times faster, from roughly 20 minutes down to 6–10 minutes; findings are described as more accurate and concise, with each one covering the problem, what triggers it, and the fix. Admins can set each repository to Quick, Standard (the default), or Deep review depth. To enable: turn Code Review on in Claude Code settings, install the GitHub App, and pick repositories; once enabled, every new pull request is reviewed automatically. Eligible Claude Enterprise organizations can claim at least $10,000 in Code Review credits by 2027-01-15; credits expire 30 days after being claimed.

Claude Code v2.1.300 — cumulative update. Added: `virtualText` in a mod's prompt decorations (text drawn in the prompt box that never enters the draft or submitted prompt); `cache_creation_5m_tokens` and `cache_creation_1h_tokens` on the OpenTelemetry `api_request` event, splitting cache writes by lifetime; OpenTelemetry GenAI semantic-convention attributes on trace spans (`gen_ai.provider.name`, `gen_ai.conversation.id`, `gen_ai.tool.name`, `error.type`, etc.); for mods, `$.ui.view({ agentId })` (shows an agent's transcript in the main view), a `dir` prop on Text for right-to-left text, and a `prompt.change` hook (raised when the prompt text changes without a typed edit), and the `agent.spawn` hook can now rewrite a workflow agent's `prompt` and `model` (previously refuse only); `/advisor` on Claude Platform on AWS (pick one with `/advisor`, `advisorModel` or `--advisor`; `CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1` turns it off); Claude Tag supports Slack Connect channels (off by default, enabled by an Owner or admin), a private Allow once / Deny approval card in Slack threads, and Slack voice clip transcripts; VS Code gains the `claudeCode.attachSelection` setting. Changed: large requests to api.anthropic.com use gzip compression (`CLAUDE_CODE_GZIP_REQUEST_BODIES=0` turns it off); exit force-kills background commands still running 1.5 seconds after being asked to stop; local session transcripts also record each API request's system prompt, tools and parameters; `updatedInput` from PreToolUse and PermissionRequest hooks is validated like Claude's own input (unknown keys dropped, rejected input denies the call); the no-op `--sandbox` and `--no-sandbox` flags were removed from `claude remote-control` (passing either now errors; set `sandbox.enabled`). Fixed: many issues with `--worktree`, reftable repositories, headless and Agent SDK session resume after compaction, `claude daemon stop`, cloud sessions and Cowork, the VS Code extension, and Code Review (GitHub).