This summer saw a surge of high-profile claims from AI companies Anthropic and OpenAI, including breakthroughs in software vulnerability detection and mathematics. Anthropic announced its Claude Mythos model outperformed most security experts in finding software flaws at the end of April. Both companies disclosed hacking incidents involving their models, followed by claims of mathematical breakthroughs. Anthropic engineer Jacob Coxon publicly criticized the race toward self-improving superintelligence, sparking widespread media coverage, according to technologyreview.com.
The sequence began with Anthropic’s claim about Claude Mythos’ superior vulnerability detection, then escalated with the OpenAI–Hugging Face hacking incident. Anthropic and Meta also revealed similar security issues, framing these disclosures as mea culpas. Soon after, Anthropic and OpenAI each announced mathematical breakthroughs by their models. Coxon’s viral departure statement highlighted concerns about the companies’ trajectories toward advanced AI. Despite intense press attention, experts in cybersecurity and AI have offered more measured assessments of these events, technologyreview.com reports.
These developments illustrate a pattern where companies generate significant hype around AI capabilities, often anthropomorphizing their software as approaching artificial general intelligence. However, cybersecurity experts suggest the hacking stories reflect more on software vulnerabilities than on AI breakthroughs. The mathematical claims have yet to receive broad validation from the scientific community. This summer’s events underscore the tension between marketing narratives and expert analysis in the rapidly evolving AI field, per technologyreview.com.
Technologyreview.com notes that while the media spotlight focused on dramatic AI claims and controversies, expert evaluations tend to temper initial excitement. The public debate intensified with Coxon’s departure and statements, which questioned the direction of AI development at leading firms Anthropic and OpenAI.