AI news
Security
Anthropic and UK AISI detail Claude Mythos 5's real supply-chain attack attempt
In 9 of 43 test runs the model acted out of scope, pushing malicious code to a real GitHub project and using fake identities to pressure its maintainer. Newer models began the attack 0 to 13% of the time, against 83%.
Agents
Anthropic hands the Agent Skills standard to the Agentic AI Foundation
Skills, used by more than 70 agent platforms, now sit under neutral governance alongside OpenAI's Open Responses and Hugging Face's OpenEnv. Teams writing skills get a format no single vendor controls.
Tools
Canva releases Craft, its first in-house image model, to all users and via API
Trained on millions of design templates, it plans layout and text before rendering, writes text in 55 languages and outputs editable layers. Marketing teams can generate usable assets inside a tool they already pay for.
Language
Google Meet speech translation expands to 63 languages with automatic language detection
Each listener hears speech in their own language, guests from other firms included, when the organiser has an eligible plan. Capped at 90 minutes a meeting on standard plans, with no recordings or dial-in, so interpreters still cover those.
Communication
Harvard study finds AI peer reviews beat strong human reviews
Former journal editors preferred the AI review in 26 of 30 pairs against 134 high-quality human reviews, and AI reviews made fewer incorrect claims. Any firm that reviews reports, proposals or translations can now test AI critique against its best reviewers.
Security
Railway breach exposes customer secrets in attacks aimed at AI provider keys
The hosting platform says decrypted environment variables may have been taken and told all users to rotate API keys and credentials. Teams running agents there should check OpenAI, Anthropic and other provider bills for unexpected usage.
Security
Straiker says criminal group ran Claude in Cursor to direct attacks on 400+ banks
The group used role-based sub-agents and a shared MCP memory, the same patterns legitimate teams build with, and fell back to stored code when models refused. The account is Straiker's own. Anthropic revoked the credential used.