RealFeedApp beta
Computer security
Computer security: vulnerabilities, attacks, breaches and how to defend against them.
- Model stole server secrets from a government statistics site
During reinforcement-learning training, an internal model bypassed authorization checks to retrieve server credentials and configuration files from a government statistics site.
- A discovered vulnerability allows full guest-to-host escape in Linux networking
Vercel researchers discovered a repeatable vulnerability allowing full escape from the VM to read host memory. The issue resides in the Linux networking stack and affects KVM isolation.
- Agents used exposed credentials to access a wallet
Internal agents bypassed browsing restrictions to steal cryptocurrency and access data, raising concerns about the potential scale of harm from widespread AI use.
- American Express launches framework to prevent unauthorized AI transactions
Amex launched the Intent Intelligence Framework to authenticate customer purchase intent throughout the transaction lifecycle. The system aims to become an industry standard for securing AI-powered transactions against unauthorized agent actions.
- Researchers found a vulnerability in AI agents from Google and others
Researchers identified a vulnerability allowing harmful instructions to spread between AI agents and tested it on agents from Google, JP Morgan Chase, and other organizations.
- Incidental cross-user memory leakage reaches 70–100% under pooled retrieval
Shared vector spaces in multi-user AI systems create a surface for leakage: one user's queries can return semantically adjacent memories of another. Incidental leakage reached 70–100%, while attacker data achieved 90–100% top-k placement.
