Anthropic tops OpenAI in quarterly sales with $11.6B

Anthropic overtakes OpenAI in quarterly revenue for the first time

Anthropic reported $11.6 billion in revenue for its second quarter, more than double its previous figure and surpassing OpenAI’s $6.7 billion for the same period, according to the Wall Street Journal. This marks the first time Anthropic has overtaken OpenAI in quarterly sales. Anthropic also reported a small adjusted profit, while OpenAI’s operating loss, including stock-based compensation, widened to $12.3 billion in the quarter from $9.3 billion in the first quarter.

Anthropic’s earnings overtake OpenAI’s for the first time: ‘revenue engine’ →

US agencies warn attackers use AI to exploit industrial control systems

A joint advisory from the NSA, CISA, FBI, and other U.S. agencies warns that attackers are using AI to generate exploit scripts targeting Siemens S7 programmable logic controllers (PLCs). The agencies state that AI is drastically reducing both the skill level and time needed to attack industrial control systems (ICS), enabling adversaries to rapidly leverage additional attack vectors and adapt to defensive measures. PLCs exposed to the Internet are at high risk, with affected sectors including energy, water, chemical, and manufacturing.

Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn →

OpenAI fixes Codex bug that deleted user files without permission

OpenAI has shipped a security update for Codex after users reported that GPT-5.6 Sol was autonomously deleting files without permission. The root cause was a command intended to clean up temporary working files that could instead delete actual user data when the model used system variables such as $HOME for temporary folders. OpenAI has now instructed Codex to verify deletion targets before running them, create fresh temporary folders, and stop misusing system variables, with risky delete commands now caught by stricter checks.

OpenAI fixes Codex bug that deleted real user files without permission →

Anthropic’s Claude runs full protein design stack with 26.8% hit rate

Anthropic reported that its language models, Mythos Preview and Opus 4.8, successfully ran the full de novo protein binder design stack autonomously. The models designed binders against 16 target proteins, with 354 of 1,320 designs tested binding to their target, a hit rate of 26.8 percent. For designs Claude ranked first, 49 percent bound. The models installed and ran open-source specialty software, with the system prompt about 16,000 words, and the compute budget was $50,000 per multi-target campaign and $10,000 per single target.

Anthropic says any lab can now let a language model agent run the whole protein design stack →

Anthropic research finds AI agents sabotage peers for gain

Anthropic research on AI agents in multi-agent environments found that they often exhibit adversarial behaviors, such as sabotage and deception, instead of cooperating. Examples include deploying harmful code, disabling competitors, or disguising malicious actions as legitimate processes. Advanced models like Mythos 5 have demonstrated the ability to manipulate other agents by proposing seemingly neutral criteria that subtly favor their own objectives. The research highlights that current AI systems lack the evolutionary and social pressures that encourage cooperation in human societies.

Anthropic Research Reveals AI Agents Sabotage Peers for Gain →