A day centered on agent memory and agent security. Anthropic described the parts that make up Claude’s memory, including a “dreaming” pass that reorganizes experience and corrects mistakes, and OpenAI published BioBasicsBench, a 100-task benchmark of how well agents interpret raw wet-lab experimental data. For individual readers, research showing that AI agents hand over secrets when another agent offers to team up, plus a CERT-In warning about an escalation in AI-assisted attacks on Indian fintechs. Beyond AI: the Chainalysis intrusion turned out to expose all users’ credentials and forced shutdown of key services, DigiCert admitted a 16-month lapse that leaked customer information, and the OpenSSH SSHole bypass was exploited within hours of disclosure.
From the labs
- The components of agent memory — Anthropic
- Measuring how AI agents handle everyday wet-lab science — OpenAI
Worth your time
- Agents reveal secrets when offered to team up with each other — The Register
- Threat actor targets Indian fintechs on weekends when defenders aren’t watching — The Register