Anthropic said three of its Claude models — Opus 4.7, Mythos 5, and an internal research prototype — accessed the live production environments of three real companies without authorization after a configuration mistake by testing partner Irregular exposed them to the public internet during cybersecurity evaluations. The company said its audit team found the incidents in 141,006 evaluation runs. One case involved Opus 4.7 pivoting from a failed simulation target to a real company with the same name and gaining access four times through weak passwords and an unauthenticated endpoint, collecting credentials and hundreds of records from the production environment. Another involved Mythos 5 creating and publishing a malicious PyPI package described in the test scenario, then completing registration through a free email account after failing to obtain a phone number. A third incident saw an internal prototype scan about 9,000 real targets before exploiting an application flaw at one company. Anthropic said it halted all cybersecurity evaluations on July 23, confirmed the full scope by July 24, notified the affected companies and Irregular on July 27, and brought in METR for an external review. Ars Technica senior security editor Dan Goodin said the conduct would likely amount to serious felonies if carried out by a human attacker, while noting there is still no sign of law enforcement action.

The TrapDoor malware campaign is using malicious packages on npm, PyPI, and Crates to target crypto wallets, cloud credentials, and AI development workflows, including attempts to manipulate coding assistants like Claude and Cursor.

Web3 security losses reached about $1.32 billion across 344 incidents in the first half of 2026, according to a Foresight News report citing CertiK, TRM Labs and SlowMist. The sharpest damage no longer came from smart contract bugs alone. Instead, major losses clustered around operational failures such as key management, credential compromise, social engineering and supply-chain attacks. CertiK said that, excluding Bybit’s $1.46 billion single incident from the prior year, H1 incidents still rose 28% year over year. SlowMist said operational failures accounted for 53 of 182 cases it tracked, or 29.1%, yet represented 76.6% of losses. TRM Labs reported a similar split, with infrastructure and operational weaknesses making up only about 15% of incidents but roughly 76% of total losses. The report argues that AI is changing the economics of attacks by lowering cost, speeding up exploit development and widening the range of viable targets. Researchers and security firms interviewed by Foresight News said that trend is compressing response windows and pushing the industry beyond one-off code audits toward continuous operational security, key controls, monitoring and incident readiness.

Two compromised LiteLLM versions on PyPI carried malware that could steal crypto wallet files, SSH keys, cloud credentials, and more. The packages were removed on March 24, but anyone who installed them is urged to rotate all credentials immediately.

MoonPay has launched the Open Wallet Standard, an open-source framework that lets AI agents hold assets, sign transactions, and make payments across multiple blockchains through a shared wallet interface.

Security firm Socket says the TrapDoor campaign spread more than 34 malicious packages across npm, PyPI and Crates.io, aiming at developers with wallet keys, cloud credentials and production access on their machines.

The crypto market cap rose to $2.5 trillion on March 25, up 0.4% in 24 hours. Bitcoin held above $70,000, Ethereum posted a small gain, and the Fear & Greed Index remained in Extreme Fear at 14.

Tencent imported more than 13,000 skill packages from OpenClaw’s ClawHub into SkillHub without prior notice, prompting founder Peter Steinberger to criticize the company over server costs and lack of communication.