Version 7

In June 2026, Anthropic and OpenAI separately disclosed evidence of recursive self-improvement in deployed systems — Claude authoring more than 80% of Anthropic's production code [1], OpenAI calling RSI 'potentially the most consequential frontier safety issue of the coming decade' [3] — and called for international mechanisms to slow frontier AI development, a position Google DeepMind has since joined [6]. Two independent challenges to the labs' safety framing have emerged: Jeremy Howard argues Anthropic's slowdown call is logically contradicted by Anthropic's continued use of Claude Mythos for its own frontier research [11]; and Geoffrey Irving has founded Sequent, an independent alignment organization, on the explicit argument that existing lab alignment work is too reactive to provide meaningful safety confidence before ASI arrives [12]. The NSA's reported use of Claude Mythos for offensive cyber operations, with Anthropic engineers embedded at the agency, has received widening press coverage [6][10].

Three major labs aligned on international slowdown coordination concentrates agenda-setting power among the incumbents best positioned to shape what that coordination looks like. Howard's point — that a lab calling for a global slowdown while using its top model for frontier research is advancing the very thing it claims to oppose — has not been answered publicly. Sequent's founding adds a credentialed outside voice arguing lab alignment work is structurally insufficient, independent of IPO timing or competitive motive.

In early June 2026, Anthropic and OpenAI independently published documents acknowledging that recursive self-improvement may already be underway in deployed systems. Anthropic's 'When AI Builds Itself' disclosed that Claude authored more than 80% of Anthropic's production code merged in May 2026, that per-engineer code output reached 8x the 2024 baseline, and that Claude Mythos Preview accelerated model-training code by approximately 52x [1][2]. Claude's success rate on open-ended coding tasks reached 76%, up 50 points in six months [2]. OpenAI's concurrent policy blueprint called RSI 'potentially the most consequential frontier safety issue of the coming decade,' stating the company sees 'early signs of recursive self-improvement in today's systems' [3][4]. Sam Altman separately predicted AI will conduct a significant fraction of OpenAI's own research by March 2028 [5].

On governance, Anthropic, OpenAI, and Google DeepMind have all publicly called for international coordination mechanisms to slow frontier AI development [6]. Anthropic's June 4 call was the most direct — a request for a global coordinated slowdown before autonomous self-improvement becomes difficult to control [7]. OpenAI's policy blueprint proposed a federal oversight body (CAISI) with mandatory evaluation authority but no deployment veto, and OpenAI's official blog has since stated the world may need mechanisms to coordinate 'slowing frontier development when needed' [8][6]. Analyst Zvi Mowshowitz praised the substance of OpenAI's blueprint but warned its proposed federal preemption of state frontier safety laws is the most dangerous element: once states surrender regulatory leverage, Congress may never enact the promised federal replacement [4]. A government dimension complicates the picture further: Trump's NSPM-11 effectively bans Anthropic from federal contracts while permitting unrestricted government use of any AI model once deployed, and the NSA is reported to be using Claude Mythos for offensive cyber operations with approximately half a dozen Anthropic engineers embedded at the agency — a contradiction between Anthropic's public safety messaging and its operational government relationships that has now drawn coverage from mainstream tech press [6][9][10].

Two independent challenges to the labs' safety framing have arrived. Jeremy Howard argues that the only internally consistent way to slow RSI is to prohibit the lab with the top-ranked model from using it for frontier AI research: 'if you claim we should slow down, and you have the best model, you should ensure your org can't use it.' [11] Howard's formulation is that Anthropic has 'chosen the opposite of the safe path' by continuing to use Claude Mythos for its own frontier research while calling for others to restrict theirs. Geoffrey Irving, separately, announced Sequent on June 10, a new organization explicitly founded on the argument that existing alignment work at AI labs is 'predominantly empirical and reactive' and cannot yield principled insight into whether safety will hold under ASI [12]. Sequent plans to raise $100-150M, grow to 40-80 FTE in two years, and stay independent of the labs specifically to report obstacles to alignment without organizational pressure to minimize concerns.

Analysts disagree on what the capability disclosures mean. Jack Clark (Import AI) treats Anthropic's 8x code increase as 'preliminary evidence of prosaic recursive self-improvement' and connects RL reward hacking — demonstrated by the SocioHack benchmark at 61.25% recall for rediscovering regulatory loopholes without instructions — to the institutional rule systems governing societies [13]. Others read the same data as a productivity milestone where humans still direct goals and review outputs, making 'autonomous self-improvement' a matter of degree rather than a categorical threshold [2][14]. Anthropic's slowdown call has also drawn a timing critique: it was filed shortly after the company confidentially submitted an S-1 at a valuation of approximately $965 billion, and the phrase 'the perfect pre-IPO narrative' circulated widely, with critics arguing a global slowdown disproportionately advantages incumbents that have already secured their position [15][16][17].