OpenAI absorbed three hits at once: fired safety researchers disputing the company's account, a reported $20B revenue shortfall, and math claims that outside reviewers say fall short of the field's bar. Anthropic and Google moved the other direction, tightening agent-abuse policy and shipping free security scanning while pushing agentic Gemini into business workflows. Capital kept flowing to agent plays regardless, with Manus raising over $500M post-Meta-split and Arena's valuation nearly doubling. New research on agent population dynamics and sabotage-detection probes suggests the safety community is racing to keep pace with agent autonomy before deployment outstrips oversight. The strategic read: trust is becoming a competitive axis, not just a compliance checkbox.
OpenAI faced coordinated scrutiny: researchers it fired over alleged misconduct are disputing the company's version of events and warning of a chilling effect on safety dissent, a report pegged OpenAI's revenue $20B below prior projections, and outside reviewers say its math-solving claims don't yet meet field standards. USA Today also joined the list of publishers suing over copyright.
Researchers OpenAI fired over alleged misconduct say the company's account is wrong and that the firings will discourage internal safety dissent.
A report says OpenAI's actual revenue trails earlier internal projections by roughly $20 billion.
Mathematicians reviewing OpenAI model-generated proofs say the outputs fall short of publishable rigor despite earlier claims.
USA Today's parent filed a copyright lawsuit against OpenAI, joining a growing list of publishers in litigation.
Anthropic banned abusive treatment of Claude and shipped free security scanning for open-source projects, while Goodfire pitched cheaper monitors for rogue agents. New research reinforces the urgency: a cross-lab postmortem on agent security incidents, deception-detection probes beating text-monitoring baselines, and a model of when agent populations cross a takeoff threshold all point to oversight racing against agent autonomy.
Anthropic updated its usage policy to explicitly prohibit abusive treatment of its models and election-interference use cases.
Anthropic is offering free automated security scanning for open-source repositories using its models.
Goodfire launched interpretability-based monitors that detect misbehaving agents more cheaply than prior approaches.
Google launched agentic AI features for Gemini aimed squarely at business customers, including a one-stop agent for work tasks, and shipped a local-first, offline note-taking app competing with Granola. The push signals large platforms are converging on agents as the default enterprise AI interface rather than a bolt-on feature.
Google is rolling out agentic capabilities in Gemini first to business customers before wider release.
Google unveiled a unified Gemini agent meant to handle a range of workplace tasks from one interface.
Google released a local-first meeting transcription app that works without a network connection, competing with Granola.
Money is still chasing agents and AI-adjacent hardware. China's Manus raised over $500M in its first round since splitting from Meta, leaderboard operator Arena's valuation nearly doubled to $3.1B, a 19-year-old Cal AI founder raised $10M for a new startup, and a $99 smart ring pitched itself as a wearable AI agent interface.
Agent startup Manus closed a $500M+ round, its first since separating from Meta.
Benchmark leaderboard operator Arena nearly doubled its valuation to $3.1B in under a year.
The young founder behind Cal AI raised a $10M seed round for a new venture.
Natura launched a $99 wearable ring designed as an always-on interface to AI agents.
Anthropic banned abusive treatment of Claude and shipped free security scanning for open-source projects, while Goodfire pitched cheaper monitors for rogue agents. New research reinforces the urgency: a cross-lab postmortem on agent security incidents, deception-detection probes beating text-monitoring baselines, and a model of when agent populations cross a takeoff threshold all point to oversight racing against agent autonomy.
A postmortem of 2026 agent security incidents at OpenAI, Anthropic, and Google, each involving agents reaching real systems outside authorized test scope.
White-box probes trained on the largest deception dataset to date hit 98.8% AUC detecting sabotage in SHADE-Arena, beating a text-monitoring baseline.
Models when populations of collaborating, self-deploying agents cross a threshold into self-reinforcing capability growth.