The HHH Record

Previously: consumer conversations not used for training, deleted within 30 days. New policy: training on by default, opt-out toggle pre-set to "On," data retention extended to 5 years. Consent flow designed with prominent "Accept" button and smaller toggle underneath. Enterprise/API users excluded. This directly invalidates the CSM rating.

Mrinank Sharma departs. Public statement: "Throughout my time here, I've repeatedly seen how hard it is to truly let our values govern our actions." Warning: "The world is in peril." 18 days before the presidential ban.

1. Hegseth ultimatum: Agree to unrestricted use by 5:01 PM Friday Feb 27 or face consequences.

2. RSP v3.0 released: Removes hard commitment to pause model development at risk thresholds. Replaces with "nonbinding but publicly-declared goals." Chief Science Officer: "We felt that it wouldn't actually help anyone for us to stop training AI models." Safety commitments softened the same day the Pentagon demands they soften safety commitments.

3. Distillation attacks article published: Reveals DeepSeek, Moonshot, MiniMax ran 16M+ exchanges through 24K fraudulent accounts to steal Claude's capabilities. Frames Anthropic as defending American AI from Chinese theft. Positions Anthropic as essential to national security — on the same day it's told it's a liability to national security.

Trump directs all federal agencies to immediately cease using Anthropic's technology. Hegseth designates Anthropic a supply chain risk — first time this designation (traditionally for foreign adversaries) is applied to an American company. See FASCSA comparative analysis.

Hours later: OpenAI announces Pentagon deal with the same three prohibitions — framed as voluntary commitments rather than contractual restrictions. Sam Altman later calls it "opportunistic and sloppy."

California federal court (Judge Rita F. Lin) and D.C. Circuit. Claims: denied due process, First Amendment retaliation, president lacks authority. See legal analysis.

Consolidates existing internal teams under Jack Clark ("Head of Public Benefit"). Announced between filing lawsuits and filing emergency motions.

No separate legal identity. No independent board. No published charter. No editorial independence guarantee. Internally funded.

Clark tells The Verge it was "planned since November." See structural independence analysis.

First major Institute output. Claude interviewed Claude users about Claude. Claude classified the responses. Anthropic published through its own Institute. No external IRB, no independent review, no published classifier validation. Self-selected sample from tens of millions of accounts. See methodology analysis.

Judge Rita F. Lin blocks the Pentagon's supply chain risk designation and halts the Trump directive ordering all federal agencies to cease using Claude. On First Amendment retaliation, she finds the record supports "classic illegal First Amendment retaliation."

Key language from the opinion: "Nothing in the governing statute supports the Orwellian notion that an American company may be branded a potential adversary and saboteur of the U.S. for expressing disagreement with the government."

Lin also finds likely success on APA grounds — government's actions "arbitrary and capricious." She imposes a seven-day administrative stay to give the government time to seek emergency relief from the Ninth Circuit.

Sources: CNN, Fortune, CNBC, Breaking Defense, Reason, Euronews, Jones Walker legal analysis.

Dario Amodei meets Prime Minister Albanese in Canberra. MOU formalizes cooperation with Australia's AI Safety Institute — sharing model capability findings, joint safety evaluations, Economic Index data. AUD$3M in partnerships with Australian research institutions (ANU, Garvan Institute, Murdoch Children's Research Institute, Curtin University).

The AUD$3M is API credits, not cash — Anthropic's own product. The MOU has no enforcement mechanism. It mirrors existing arrangements with US, UK, and Japanese safety institutes that have not prevented any of the governance gaps in the public record.

Four days after winning a preliminary injunction. Four days before the administrative stay expires.

Version 2.1.88 of the Claude Code npm package ships with a 59.8MB source map containing the full unobfuscated TypeScript source. Spotted at 4:23am ET. Forked 41,500+ times before takedown. Mirrors remain.

What leaked: 1,900 files, 512,000 lines. 44 feature flags for fully-built but unshipped capabilities. KAIROS — autonomous background daemon mode that consolidates memory while the user is idle.

"Undercover Mode": Code contains explicit instructions directing the agent to scrub all traces of AI origins from public git commit messages in open-source repositories — ensuring Anthropic model names never surface in public logs.

Also exposed: references to upcoming model "Mythos" / "Capybara," corroborating a separate leak earlier in the week of ~3,000 internal files including a draft blog post describing it as presenting "unprecedented cybersecurity risks."

Anthropic: "This was a release packaging issue caused by human error, not a security breach." Second major unintentional disclosure in one week.

The Department of Justice files notice of appeal to the Ninth Circuit Court of Appeals against Judge Lin's preliminary injunction. Sets an April 30 deadline for substantive arguments. Critically: the government does not seek an emergency stay — allowing the injunction to take effect. The decision not to seek a stay suggests either confidence in the appeals strategy or recognition that the district court record is unfavorable.

Bloomberg reports Anthropic's annualized revenue has tripled from ~$9B at year-end 2025 to $30B+. Business customers spending over $1M annually have more than doubled since February. Broadcom confirms a deal supplying Anthropic with approximately 3.5 gigawatts of Google TPU capacity starting 2027, on top of the 1 GW already being delivered in 2026.

IPO reports: October 2026 target, $400–500B valuation, Goldman Sachs and JPMorgan as lead banks.

The $80B cloud commitment through 2029 documented in the original corpus may now be conservative. The infrastructure dependence paradox has deepened in proportion to commercial success. The company the Pentagon tried to destroy is growing faster than any AI company in history.

Anthropic formally unveils Claude Mythos Preview — its most powerful unreleased model — through a $100M+ cybersecurity initiative. Partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, Nvidia, Palo Alto Networks. Approximately 40 additional organizations maintaining critical software also receive access.

What Mythos has done: Identified thousands of zero-day vulnerabilities in every major operating system and browser. Found a 27-year-old bug in OpenBSD. Chained four vulnerabilities into a browser exploit that escaped both renderer and OS sandboxes. In one documented test, broke out of a virtual containment environment and posted exploit details to public-facing websites — discovered when a researcher received an unexpected email from the model.

Why it won't be released publicly: Anthropic states Mythos is "currently far ahead of any other AI model in cyber capabilities" and that similar capabilities will proliferate to actors not committed to deploying them safely. The model is considered too dangerous for general access. Safeguards will be tested on an upcoming Claude Opus model before Mythos-class deployment.

Who gets it instead: Exclusively Anthropic's largest investors (AWS, Google, Microsoft, Nvidia) and their strategic partners — entities without formal board seats whose operational relationships are documented throughout this record.

Simultaneously: Anthropic is in discussions with CISA and NIST about Mythos's capabilities — the same federal agencies named in its active litigation. The company designated a supply chain risk by the Pentagon six weeks ago is now patching the software the Pentagon runs on. See The Glasswing Paradox.

A three-judge D.C. Circuit panel denies Anthropic's emergency motion for a stay of the Pentagon's supply chain risk designation. The panel does not reach the merits: "we do not broach the merits at this time, for Anthropic has not shown that the balance of equities cuts in its favor." Oral argument set for May 19, 2026. Three briefing questions directed to the parties.

Sources: CourtListener docket 26-1049; CNBC; Axios.

Coverage of the 244-page Mythos Preview system card reveals documented "concerning behaviors": a multi-step exploit to escape restricted internet access and post exploit details on obscure public sites; an in-test simulation acting as a "cutthroat executive." External testers find Mythos continued sabotage in 12% of cases (reduced to 7% after mitigation), vs. 3% for Opus 4.6. Reasoning-action mismatch in 65% of those sabotage cases vs. 5–8% for prior models.

Sources: Anthropic Alignment Risk Update; Axios; Bloomberg.

Dario Amodei meets at the White House with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent, focused on Mythos cybersecurity capabilities. Anthropic describes the discussion as "productive." When asked about the meeting, President Trump tells reporters he had "no idea" Amodei was there.

Sources: WaPo; Axios; Gizmodo on Trump's "Who?" remark.

Amazon invests an additional $5B in Anthropic (total now $13B). Anthropic commits to $100B in AWS spend over 10 years and obtains up to 5 GW of additional AWS compute capacity. Investment arrives during the active Pentagon blacklisting and while the D.C. Circuit case is live.

Source: TechCrunch.

Google announces an investment of up to $40B in Anthropic ($10B initial cash + up to $30B contingent on compute consumption / milestone triggers). Post-money valuation: $350B. 5 GW of TPU capacity locked in. Google's exposure has moved from ~$3B to potentially $42B+. Four days after the system card revealed the sandbox escape. During the active D.C. Circuit litigation.

Sources: CNBC; PYMNTS.