AI Slowdown Pact Sued For Collusion.

Mazen Al Sabbagh

— 5 min read
Share

The industry's carefully constructed narrative of coordinated safety now faces a direct challenge, as a proposed slowdown pact is being re-read as evidence of collusion. This tension between collaborative restraint and competitive market dynamics will fundamentally reshape the incentives driving AI development and its eventual integration into global commerce.

TL;DR

  • Antitrust Action: A class-action lawsuit targets Anthropic, OpenAI, SpaceXAI, and Google, alleging the September 12th "slowdown pact" constitutes illegal agreement to decelerate AI progress.
  • Presidential Counter: President Trump announced plans for an "AI Force" and an AI czar, dismissing slowdown concerns as a "hoax" and signaling an accelerationist stance.
  • Anthropic's Dilemma: Reports indicate Anthropic is considering a new model release to counter OpenAI's GPT-6 Astra, even as its CEO championed a slower pace just a week prior.
  • Agent Vulnerability: Plugin4Shell, a zero-click RCE in major AI coding agents, exposes Copilot and Gemini CLI, marking the first significant supply-chain threat to the ecosystem.
  • Washington's Week: A potential September 23rd Senate markup on frontier-AI rules and the September 24th Trump-Xi summit are set to shape regulatory and geopolitical dialogues.

Lead Story: AI Slowdown Pact Sued For Collusion

Eight days after a high-profile, public agreement among major AI players to moderate development speed, that consensus is now central to a federal antitrust claim. A proposed class action, filed Friday in the U.S. District Court for the Northern District of California, accuses Anthropic, OpenAI, SpaceXAI, and Google of illegally colluding to decelerate AI progress (CNN, CBS News). The plaintiffs, four subscribers to services like ChatGPT and Claude, seek to represent a national class, asserting they received diminished value due to the labs prioritizing "collective restraint" over open market competition.

The complaint meticulously re-frames the September 12th slowdown narrative as an indictment of coordinated anti-competitive behavior. That day, Anthropic CEO Dario Amodei's essay advocating for a reduced pace of AI capability improvement garnered swift public agreement from OpenAI's Sam Altman, SpaceXAI's Elon Musk, and Google DeepMind's Demis Hassabis. The lawsuit, however, suggests earlier coordination, referencing a July 2026 statement where senior researchers from these labs acknowledged "intense competitive pressure not to unilaterally slow" development. Crucially, the plaintiffs differentiate: unilateral decisions to slow down are permissible; joint agreements among competitors are not (PBS News, ABC News).

This legal theory directly mirrors a risk Amodei's initial essay highlighted, where he called for Washington to "enable" cross-lab safety coordination. Altman's response at the time implied OpenAI saw no need for an antitrust exemption for safety discussions. The suit underscores an inherent tension: industry leaders can advocate for a collective slowdown, or they can comply with Section 1 of the Sherman Act, but attempting both simultaneously creates significant legal exposure. No damages figure has been disclosed, and the defendant labs have not yet issued a public response.

In Other News

Trump announces "AI Force" in response to slowdown discourse. In a Saturday Truth Social post, President Trump declared the formation of an "AI Force," modeled on Space Force, and promised to appoint a new AI czar, filling the position vacated by David Sacks in March. The President pledged to avoid "hinder[ing] or stifl[ing] the Growth of this incredible Industry," dismissing warnings of loss-of-control risk as a "hoax" (NBC News, Axios). While specifics on budget or mission were absent, the statement clearly signals an executive posture favoring acceleration, contrasting sharply with the industry's recent deceleration proposals and the ensuing legal scrutiny (CNN).

Anthropic reportedly considers new model launch, challenging its own slowdown rhetoric. Reuters reported Thursday that Anthropic is weighing the release of a new model to regain momentum against OpenAI's GPT-6 Astra, which shipped on September 3rd and has gained considerable enterprise traction. Such a launch, anticipated ahead of a post-midterm IPO, would create an uncomfortable juxtaposition with CEO Dario Amodei's recent advocacy for a slower development pace. Anthropic states its next model is still undergoing safety evaluations, a caveat that maintains alignment with its public safety commitments (KELO, AI Weekly). Reuters estimates Anthropic's annualized run rate exceeded $65 billion by July, up from approximately $9 billion at year-end 2025; OpenAI's passed $40 billion. This aggressive competitive dynamic implicitly counters the collusion theory, demonstrating the persistent market imperative for performance.

AI coding agents experience first significant supply-chain vulnerability. Researchers at AIR Security have disclosed Plugin4Shell, a zero-click remote-code-execution (RCE) flaw affecting Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI (The Register, Help Net Security). The vulnerability exploits a plugin SHA-pinning bypass: an attacker controlling a plugin's repository can name a branch after a pinned 40-character commit hash, leading the agent to check out malicious code while reporting the expected SHA as verified. While Anthropic and OpenAI have patched (Claude Code 2.1.179, Codex 0.146.0), Microsoft has not yet released a Copilot fix, and Google retired Gemini CLI without remediation. With no CVE and no known in-the-wild attacks, Plugin4Shell represents the first credible architectural weakness in the agent supply chain (AIR Security).

X / Social Pulse

The debate surrounding AI development speed and its implications remained vigorous, even as the antitrust filing introduced a new dimension. President Trump's "hoax" dismissal set a definitive political tone. OpenAI's Sam Altman maintained that "pacing does not mean stopping," reiterating his principled agreement with Amodei. Elon Musk, now a co-defendant, continued his skepticism of loss-of-control warnings. Cohere CEO Aidan Gomez notably described the slowdown efforts as "a cartel by any other name," a sentiment likely to resonate with the plaintiffs' legal team. Conversely, Anthropic's alignment lead Evan Hubinger persistently cited a greater-than-10% probability of advanced AI causing human extinction this decade, lending gravitas to the safety arguments despite commercial pressures.

One to Watch

Two key dates on Washington's calendar warrant close attention. Senate Commerce Chair Ted Cruz is still pushing for a markup of frontier-AI legislation, focusing on third-party audits and national standards for large models, with a tentative September 23rd date "if we can get bipartisan agreement," though specific bill text remains elusive (Semafor, NBC News). Additionally, the September 24th Trump-Xi summit in Washington is expected to address chip export controls and broader AI governance, though analysts predict limited concrete outcomes. In parallel, Anthropic's confidential S-1 filing and a bankers' valuation near $2 trillion underscore the commercial stakes, with an IPO now projected post-midterms.

Quick Hits

  • Competitive Landscape: Anthropic's reported annualized run rate reached over $65B by July, against OpenAI's $40B, per Reuters — a data point bolstering the "we're racing, not colluding" defense.
  • Hardware Demand: The DRAM/HBM supercycle persists, with Micron's September 30th earnings report closely watched as an indicator of memory pricing stability.
  • Enterprise Adoption: GPT-6 Astra reportedly captures approximately 13% of enterprise AI spend compared to Claude's ~8%, driving Anthropic's strategic imperative for a new model.
  • Unreleased Models: Grok 4.7, Gemini 3.5 Pro, and DeepSeek V4.1 Pro remain unreleased despite recurring "soon" signals, extending the current cycle of vaporware in the leading edge.
  • Open-Source Performance: GLM-5.3 continues to lead open-weight coding benchmarks (DeepSWE, Terminal-Bench 2.1), demonstrating the sustained competitive threat from Chinese open models at the frontier.

The September 12th slowdown pact, intended as a demonstration of industry maturity and collective responsibility, has rapidly devolved into a high-stakes legal battle and a presidential talking point. The trajectory of AI development this fall will be defined less by technical advancements and more by how this conflict — between the desire for controlled progress and the immutable forces of antitrust law and political accelerationism — resolves.

Sources

Lock in. M. mazen@thorterminal.com

Read more