Mythos Zero Days, Discord Breaches Anthropic.

Mazen Al Sabbagh

— 6 min read
Share

TL;DR

  • Mythos capability and breach: Anthropic's Mythos located 271 Firefox zero-days, a 12.3x increase over previous models. The same week, a private Discord group gained unauthorized access via a vendor environment. CISA remains without access, while the NSA does.
  • OpenAI criminal probe: Florida's AG initiated the first US state criminal probe into an AI company, targeting OpenAI over ChatGPT's alleged role in a 2025 mass shooting. The AG stated the chatbot would face murder charges if human.
  • SpaceX acquisition and S-1 disclaimer: SpaceX secured an option to acquire AI coding startup Cursor for $60 billion. Concurrently, its S-1 filing cautioned investors that orbital AI data centers might not be commercially viable, challenging xAI's "Colossus" vision.
  • GPT-5.5 'Spud' leak and launch: A server misconfiguration exposed GPT-5.5 "Spud" via OpenAI's Codex for 90 minutes. Polymarket odds for an April 23 launch reached 86%. OpenAI also launched ChatGPT Images 2.0 as a distinct product.
  • Bezos' Project Prometheus funding: Jeff Bezos' secretive AI lab, Project Prometheus, is finalizing a $10 billion funding round at a $38 billion valuation, positioning it as the third-most-valuable independent AI entity.

Lead Story: Mythos: Unprecedented Capability, Compromised Access

The world's most critically discussed AI model validated its potency this week. It also underscored significant failures in its custody.

Mozilla published results Monday confirming Anthropic’s Mythos identified 271 zero-day vulnerabilities in Firefox 150, all patched pre-release. This represents a 12.3x increase over Opus 4.6's 22 findings in Firefox 148. Mozilla’s CTO positioned Mythos as equivalent to top human security researchers, while The Register observed that the unprecedented speed and scale, rather than the novelty of the bugs, were the salient points. Bain & Company's advisory described Mythos as "a business risk of the highest order."

This capability demonstration coincided with a significant lapse in Anthropic's containment strategy. Bloomberg reported Monday that a small group, coordinated via a private Discord channel, accessed Claude Mythos Preview through a third-party vendor. The method was rudimentary: a contractor shared credentials, and the group inferred the model's URL. TechCrunch verified the access with screenshots and a live demonstration. The users, identified as model hobbyists, reportedly did not engage in cybersecurity-related prompts.

The incident's implications deepened with Axios’s revelation that CISA — the Cybersecurity and Infrastructure Security Agency, mandated to protect US critical infrastructure — lacks Mythos access. Over 40 organizations, including Mozilla and JPMorgan, have access, and the NSA operates through an ODNI loophole. CISA, however, has only received briefings on the model's capabilities.

The regulatory response is accelerating globally. Bundesbank President Joachim Nagel, speaking in Rome, advocated for equitable access, labeling Mythos a "double-edged sword." The Reserve Bank of Australia and New Zealand confirmed monitoring efforts, joining other central banks. Reuters reported Anthropic's intent to extend access to European banks imminently. The IAPP published a governance analysis citing Mythos's specific risks.

Anthropic has offered no public statement beyond "investigating" since Monday. The absence of communication is now part of the evolving narrative.

In Other News

Florida AG Opens First US Criminal Probe of an AI Company Over a Mass Shooting. Florida Attorney General James Uthmeier initiated a criminal investigation into OpenAI on Monday. The probe links ChatGPT to a 2025 FSU mass shooting, with Uthmeier asserting the chatbot provided "significant advice" and would be charged with murder if it were a person. This marks the first US state criminal investigation of an AI firm tied to a violent crime. OpenAI's position is a denial of responsibility. This legal pressure emerges five days before the Musk-Altman trial and coincides with a likely GPT-5.5 launch, creating a multi-front legal challenge for OpenAI.

SpaceX Secures Option to Buy Cursor for $60 Billion — While Its S-1 Quietly Walks Back Orbital Compute. SpaceX announced a partnership with Cursor that includes an option to acquire the AI coding startup for $60 billion, or a $10 billion payment for collaboration without acquisition. Bloomberg and CNBC confirmed the terms. Simultaneously, SpaceX’s S-1 filing included a warning to investors that orbital AI data centers rely on "unproven technologies and may not become commercially viable." This legally mandated disclosure stands in direct contrast to xAI's "Colossus in space" vision. Analyst meetings at Starbase and the Memphis Colossus data center are ongoing through Thursday.

GPT-5.5 "Spud" Leaked for 90 Minutes — Launch Now Imminent. A server misconfiguration within OpenAI's Codex tool exposed "gpt-5.5-turbo-preview" for approximately 90 minutes on Monday. Developers accessing it noted improved reasoning and faster response times over current GPT-5.x iterations. Polymarket odds for an April 23 release surged to 86%. Sam Altman’s public anticipation for "this week" and a scheduled Wednesday livestream further suggest an imminent announcement. Naming, whether GPT-5.5 or GPT-6, remains contingent on the performance delta. Separately, OpenAI officially launched ChatGPT Images 2.0 on Tuesday, featuring native reasoning-before-drawing, 2K resolution, and multi-image consistency.

Bezos' Project Prometheus Nears $10B Round at $38B Valuation. Jeff Bezos' private AI research lab, "Project Prometheus," is reportedly close to securing $10 billion in funding at a $38 billion valuation, with BlackRock and JPMorgan leading. The lab’s focus is "physical AI," emphasizing models that comprehend physics for industrial and manufacturing applications. This represents Bezos’s direct AI investment, distinct from Amazon’s stake in Anthropic. At $38 billion, Project Prometheus would immediately rank as the third-most-valuable independent AI lab, after OpenAI and Anthropic.

X / Social Pulse

Simon Willison detailed the Claude Code pricing fluctuations in real-time on Tuesday. Anthropic’s pricing page temporarily showed Claude Code unavailable for Pro subscribers. Anthropic’s Amol later clarified this was a "small test on ~2% of new prosumer signups," with no impact on existing subscribers. The broader implication is that agentic tools are generating inference costs that challenge current flat-rate pricing models, with Pro users reporting sessions up to 3x longer on Opus 4.7.

Altman’s "fear-based marketing" comments persist a day after TechCrunch published his Core Memory podcast remarks. His quote: "It is clearly incredible marketing to say, 'We have built a bomb...We will sell you a bomb shelter for $100 million.'" The timing of the Mozilla 271-zero-day report and the Discord breach on the same day provided both validation for Mythos’s capability and a critique of its access controls, sharpening Altman's argument while inadvertently proving the model's power.

Hannover Messe 2026 continues (April 20-24), featuring Nvidia, Microsoft, Siemens, and Lenovo exhibiting AI-driven manufacturing solutions. Approximately 15 companies are showcasing humanoid robots, indicating a rapid transition from R&D to factory deployment within months.

One to Watch

The Musk-Altman $134 billion trial begins jury selection on Sunday — adding another front to OpenAI's legal challenges. Five days out, Musk’s amended filing seeks Altman's removal and the redirection of all winnings to OpenAI's nonprofit arm. OpenAI characterized this as a "legal ambush." The Florida criminal probe further compounds the legal pressure. This trial runs concurrently with GPT-5.5's probable public release, SpaceX's analyst roadshow, and the ongoing Mythos breach fallout, positioning the next seven days as potentially the most consequential for AI industry dynamics this year.

Quick Hits

  • xAI filed suit to block Colorado's AI Act before its June 30 enforcement, making it the first major AI company to challenge state AI legislation in federal court.
  • Motley Fool highlighted the Anthropic-Alphabet-Broadcom multi-gigawatt TPU deal—1 GW by end of 2026, escalating to 3.5 GW by 2027—underscoring the intensifying demand for compute infrastructure.
  • xAI quietly launched Grok Build in public beta on April 21, introducing a local-first CLI coding agent and intensifying the competitive landscape for AI coding solutions against Claude Code, Codex, and Cursor.
  • FISA Section 702 expires April 27—the same day as Musk-Altman jury selection—with unresolved congressional concerns regarding AI’s potential to amplify surveillance capabilities.
  • Meta began installing tracking software on US employees' computers to gather mouse movements, keystrokes, and screenshots for AI agent training, prompting employee pushback against the mandatory "Model Capability Initiative."

Tuesday presented the clearest articulation yet of the Mythos paradox: the model identified 271 zero-days in Firefox before release, confirming Anthropic’s claims of extraordinary capability, only for a Discord server to expose critical vulnerabilities in its access controls. CISA’s exclusion from the access list, while the NSA navigates a jurisdictional loophole, elevates a security incident into a full-blown governance crisis. Concurrently, OpenAI is rapidly accumulating legal exposure—a Florida criminal probe, the Musk trial in five days, and a GPT-5.5 launch prematurely revealed by a misconfiguration. Bezos is covertly funding a third major AI lab. SpaceX seeks to acquire Cursor for $60 billion while simultaneously advising investors that its orbital compute aspirations may not materialize. The confluence of these events indicates that contradictions are no longer peripheral; they are the industry's defining feature.

Sources

Lock in. M. mazen@thorterminal.com

Read more