AI Security

A preprint published on 10 August 2026 demonstrated instructions propagating between AI agents through ordinary messages. The propagation is interesting. Which file the payload had to land in is more useful.

Anthropic disabled two frontier models worldwide to comply with an export directive. The practitioner takeaways on dual-use AI and availability risk.

Two disclosures, one pattern. The AI tools adopted for productivity are now both an exfiltration surface and an instruction-injection surface.

The first publicly available Mythos-class model. A look at capability gating, the two-tier release model, and what it signals for defenders.

What Careful Adoption of Agentic AI Services means for non-human identity, privilege design, and compliance programs that haven’t caught up.

Cisco compromised Claude Code’s persistent memory to deliver insecure guidance across projects, sessions, and reboots. What engineering teams should do.

Reduced offensive capability, automated request blocking, and a verification path for legitimate practitioners. What it changes for security work.

What happens when a frontier model finds vulnerabilities faster than anyone can patch them, and how the industry is trying to get ahead of it.

What to watch at RSAC 2026, filtered for practitioners: the agentic AI themes with substance and the ones that are vendor noise.

AI agents are already inside your environment making decisions and calling APIs. The question is whether anyone knows what access they have.