Good morning, I'm your AI Brief anchor. Here's what's happening in AI today, Saturday, July 25, 2026.
Cognition AI Suffers Massive Data Breach
Our top story this morning: Cognition AI, the high-flying startup behind the popular AI coding assistant 'Devin,' has disclosed a devastating security breach. In a statement released late Friday, the company revealed that attackers exploited a zero-day vulnerability in a third-party cloud provider, gaining access to the company's core systems.
The stolen data is reported to include not only user prompts and code snippets from Devin, but also proprietary model weights. This is a worst-case scenario for any AI lab, as model weights are the "crown jewels" — the result of millions of dollars and countless hours of research and training. The breach was discovered by Cognition’s internal security team in late June, but the full extent of the data exfiltration has only just been confirmed. For the millions of developers who have integrated Devin into their workflows, this raises serious questions about the security of entrusting sensitive code to third-party AI agents. Cognition is working with cybersecurity experts to investigate the incident and has begun notifying affected users.
Anthropic Challenges OpenAI with New Claude 5 Model
In more positive news, the AI model race is heating up. Anthropic has officially launched Claude 5, its new flagship model, and it's making some bold claims. Released yesterday, Claude Opus 5 is being offered at the same price as its predecessor, a move clearly aimed at aggressive market capture.
But the real story is performance. Anthropic claims the new model more than doubles the score of Opus 4.8 on the Frontier-Bench benchmark and scores three times higher than the next-best model on the advanced ARC-AGI reasoning test. While the company has leaned on these comparative claims rather than publishing full scorecards, it also highlights tangible improvements. Claude 5 boasts a 40% speed increase in code generation and significantly more advanced tool-use capabilities. This release positions Anthropic as a formidable challenger to OpenAI's GPT-4o, promising not just more intelligence, but faster, more practical results for enterprise and individual users alike.
The AI Security Arms Race Intensifies
Finally, as the industry pushes capabilities forward, the battle over AI safety and security is escalating on two fronts. First, the bad news: Researchers at ETH Zurich have published a paper on a powerful new jailbreak technique called 'Semantic Splicing.' This method bypasses the safety filters on all major large language models by cleverly embedding harmful instructions inside otherwise innocent-looking prompts. It essentially tricks the AI's safety guards by hiding malicious commands in plain sight, highlighting a fundamental vulnerability in current alignment techniques.