This week's articles
Defense in Depth, Rebuilt for Machine Speed
Cheap AI-built exploits now outpace patching, so a16z say defenses must work without fast patches: remove exposure, segment networks, block attacks at the network layer, eliminate bug classes, and prove backups restore. They outline six opportunities, including live fixes, rule-free detection and a long-tail platform.
Sandboxing and AI
This blog post is an exploration of AI sandboxes, starting with the requirements we want from a sandbox, and an overview of the various underlying technologies available, and then some specific product/services in the market.
Scaling Security Code Reviews with an Agentic Prover
1Password added a read-only, sandboxed agentic "Prover" stage to SAGE. It checks findings against the wider codebase. On a 125-finding test it suppressed 85 of 86 false positives and kept all four real, unprotected flaws.
Things LLMs get wrong about AWS IAM
Agents writing AWS IAM have picked up weird habits: embellishing with Deny in identity policies, using permissions boundaries as a generic restriction, and adding sts:GetCallerIdentity in policy documents.
On the hunt for ARTEX: understanding agentic attacks in AWS environments
Datadog reports in-the-wild AWS activity from ARTEX, an LLM-agent penetration framework. Vectors: compromised long-term IAM keys, used for CloudWatch FilterLogEvents searches for financial data and secrets; and anonymous public Cognito user pool sign-ups, yielding tokens and credential enumeration.
|