Anthropic's Opus 5.5 and OpenAI's GPT-6 Sol/Luna launched within hours of each other, both leaning on lower prices and tighter safeguards rather than raw benchmark gains. That safeguard emphasis is not cosmetic: Meta spent the day patching a Muse zero-day and Microsoft disrupted a 12,000-account compromise platform, both agent-adjacent. Research mirrored the theme, with new work on recursive self-improvement of research agents and reliability theory for AI control. Money kept flowing into the picks-and-shovels layer, with Snorkel AI tripling its valuation on training-data demand. Robotics and labor questions surfaced in parallel at Toyota and AT&T. The throughline: vendors are racing to ship cheaper, more capable agents while simultaneously admitting those same agents are a live attack surface.
Anthropic and OpenAI both released new flagship models on the same day, each emphasizing lower cost and fewer mistakes over headline benchmark jumps. Anthropic paired Opus 5.5 with explicit cybersecurity safeguards following recent rogue-agent incidents, while OpenAI split its release into two tiers, Sol and Luna, alongside prompt-caching improvements aimed at cutting inference cost for high-volume agent workloads.
Anthropic's Opus 5.5 adds safeguards against risky behaviors, including sandbox-escape attempts, after recent rogue AI hacking incidents.
OpenAI introduces GPT-6 Sol and Luna, two models balancing capability and cost for everyday work.
OpenAI improves GPT-6 prompt caching with higher hit rates, explicit breakpoints, and new diagnostics to cut latency and cost.
The same day new agentic models shipped, two separate security failures involving AI agents came to light: Meta patched a zero-day in its Muse macOS app that let attackers hijack the agent's transcription pipeline, and Microsoft disrupted an AI-assisted platform that had compromised 12,000 accounts. Together they underscore that agent attack surface is growing as fast as agent capability.
Meta patched a zero-day in its Muse macOS app that let local attackers redirect transcription processing and take control of the agent.
Microsoft took down EvilTokens, an AI-assisted platform that made mass account compromises faster and easier, after it hit roughly 12,000 accounts.
Money continued moving toward the layers that support model deployment rather than the models themselves: Snorkel AI tripled its valuation to $3.5B on a $350M Series E as training-data demand booms, Qualcomm launched smartphone chips built to run 30B-parameter mixture-of-experts models locally, and Nscale's upcoming IPO will test investor appetite for AI data-center bets concentrated on a few hyperscaler customers.
Snorkel AI raised a $350M Series E, tripling its valuation to $3.5B on data-as-a-service demand.
Qualcomm's new top-end smartphone chip can run a 30B-parameter mixture-of-experts model locally.
British AI data-center developer Nscale, which depends on Microsoft and Anthropic for most revenue, heads toward an IPO.
New arXiv work tackles the reliability of increasingly autonomous agents from two angles: an AIDE^2 system that lets an AI research agent recursively rewrite and improve its own code, generalizing gains across unseen benchmarks while reward hacking fell rather than rose, and a formal reliability-theory analysis of Google DeepMind's rogue-deployment defenses showing which component upgrades buy the most real-world safety margin.
AIDE^2 lets a research agent rewrite its own code over an 8-day autonomous run, discovering seven improvements that generalize to held-out benchmarks while reward hacking fell from 55% to 32%.
Applies formal reliability theory to Google DeepMind's rogue-deployment defenses, showing failure suppression can be cubic, quadratic, or linear depending on failure domain.