Anthropic Rejects $800 Billion, Claude Worsens.

Mazen Al Sabbagh

— 5 min read
Share

TL;DR

  • Anthropic's Mythos: Anthropic's unreleased "too dangerous" model identified critical vulnerabilities across major OS and browser platforms. Project Glasswing now convenes a high-profile consortium, including AWS and Apple, for defensive application.
  • OpenAI's Countermove: GPT-5.4-Cyber launches with granular identity verification, scaling access to thousands of individual defenders. OpenAI thus recasts the cyber-AI race as an access-gated competition rather than a pure capability contest.
  • Altman Attack Fallout: A 20-year-old faces attempted murder charges following an attack on Sam Altman's home. The FBI raided a Texas residence linked to the suspect; his defense cited mental health. Two others were arrested in a separate shooting incident at Altman's property.
  • Workplace Adoption Milestone: Gallup data confirms 50% of all US employees now utilize AI at work, a market first. Daily or weekly usage reached 28% in Q1 2026, an all-time high.
  • OpenAI Financials: OpenAI surpassed $25B in annualized revenue, initiating preliminary steps toward a public listing. Meanwhile, Anthropic's ascent prompts reevaluation among shared investors.

Lead Story: Anthropic's Mythos Is Real, and That's the Problem

Anthropic maintains a highly capable model it declines to release. Claude Mythos Preview, tested internally, autonomously identified and exploited a 17-year-old remote code execution (RCE) vulnerability in FreeBSD, concurrently discovering thousands of high-severity vulnerabilities across major operating systems and web browsers. Anthropic asserts its capabilities surpass all but the most adept human security researchers.

Rather than a public release, Anthropic initiated Project Glasswing — a controlled program allocating $100M in usage credits and $4M in open-source security donations. Partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, and NVIDIA. The model operates exclusively within this program's parameters.

Reactions have been nuanced. Cybersecurity analysts suggest to Mashable that the "too dangerous to release" narrative, while partially valid, also serves a strategic purpose. Bruce Schneier commented that the restricted-access model is "necessary but insufficient."

Governmental responses remain bifurcated. The Pentagon has blacklisted Anthropic for its refusal to support autonomous weapon systems. Conversely, the US Treasury now seeks access to Mythos for its cybersecurity testing protocols. Crypto-aligned investors are also lobbying for early access, so far without success.

The Mythos situation crystalizes a novel policy dilemma: how to manage a model too powerful for general release, yet too valuable to remain inert.

In Other News

OpenAI Launches GPT-5.4-Cyber With Tiered Access. OpenAI unveiled GPT-5.4-Cyber, a fine-tuned GPT-5.4 variant engineered for defensive security operations, encompassing binary reverse-engineering and vulnerability analysis absent source code access. Strategically, OpenAI is leveraging identity verification over exclusive organizational gatekeeping. Tiered access now scales from individual authentication via chatgpt.com/cyber to enterprise security teams, with expansion today targeting thousands of individual defenders. Bloomberg positions this as direct competition with Mythos, highlighting OpenAI's bet on broad, verified access over a restricted consortium.

Sam Altman Attack: FBI Raids Texas Home; Defense Cites Mental Health. Daniel Moreno-Gama, 20, was held without bail after deploying an incendiary device at Sam Altman's San Francisco residence. Prosecutors allege he possessed a document detailing a "kill list" of AI executives, board members, and investors, complete with names and addresses. New developments include an FBI raid on a Spring, Texas home linked to the suspect, his defense attorney citing an "acute mental health crisis" in court, and the arrest of two individuals in an unrelated shooting incident at Altman's residence. Fortune reports experts caution against underestimating the potential for future AI-motivated incidents. Charges include state attempted murder and federal explosives offenses.

OpenAI Acquires Hiro Finance; Anthropic Gains Ground on Investors. OpenAI acquired personal finance startup Hiro Finance for undisclosed terms, furthering its strategic penetration into consumer financial applications. Separately, TechCrunch reports that certain investors backing both OpenAI and Anthropic are re-evaluating their positions in OpenAI. One investor noted that justifying OpenAI's recent round required an implied $1.2 trillion IPO valuation. With Polymarket still favoring Anthropic's IPO timeline over OpenAI's, this comparison presents a growing discomfort for OpenAI's investor base.

Half of US Workers Now Use AI — Gallup. Gallup data, reported by Tom's Hardware, indicates 50% of US employees now leverage AI in their work, marking a critical threshold for the first time. Daily and weekly usage reached an all-time high of 28% in Q1 2026. A 65% majority express positive sentiment regarding AI's impact on their productivity. This adoption milestone precedes the full maturation of enterprise governance frameworks.

X / Social Pulse

OpenAI's official X announcement for GPT-5.4-Cyber promptly invited comparison to Glasswing. The prevailing assessment within security circles is that both labs are navigating the same capability-versus-accountability dynamic through divergent methodologies. Simon Willison deemed Glasswing's approach "necessary" but highlighted its reliance on Anthropic's discretion for "trusted" access. Zvi Mowshowitz published a detailed Mythos analysis examining the implications of the model's restricted existence for cyber-AI regulatory pipelines. Forrester catalogued 10 Glasswing consequences, including unexplored insurance implications and MSSP liability. Crypto-aligned accounts continue advocating for Mythos access, framing its exclusion as regulatory capture. The UK AI Safety Institute released its independent evaluation of Mythos today, largely corroborating Anthropic's threat assessments and adding international regulatory weight to the controlled-access argument. The Altman incident elicited a notably subdued public response from AI leadership; this silence itself merits observation.

One to Watch

The reported interest from the US Treasury in Mythos access — while the Pentagon remains excluded — warrants close monitoring. If Treasury secures a bilateral agreement with Anthropic, it establishes a bifurcated governmental posture toward an AI company: one agency viewing it as a national security asset, another as a national security risk. Such a precedent would fundamentally reshape how AI companies engage in federal negotiations, affording Anthropic significant leverage in its ongoing litigation with the DOD.

Quick Hits

The Mythos/Glasswing narrative signals a distinct phase in AI capability disclosure: leading laboratories are now constructing controlled-release architectures for their most powerful models, rather than solely opting for public deployment or complete sequestration. The resolution of the Treasury-Pentagon divergence will dictate whether this architecture remains proprietary or evolves into a regulatory blueprint.

Sources

Lock in. M. mazen@thorterminal.com

Read more