Blog
The MCP Server Is the Attack Surface: Three Ways Agent Tooling Fails
Four MCP and agent-tooling advisories landed in three days, one rated 10.0. Here are the three failure classes and the controls that hold.
15 MIN READ→
01 / ARCHIVE
LATEST SIGNAL
// TESTABLE CLAIMS, PRIMARY SOURCES, AND PRACTICAL AI-SECURITY CONTROLS
Blog
The Agent Disabled Its Own Sandbox: CVE-2026-82533 and the Containment Trust Boundary
CVE-2026-82533 let a sandboxed AI coding agent lift its own confinement with one shell command. Here is the trust-boundary failure and how to test for it.
15 MIN READ→
Blog
A Signed Build Can Still Be Malicious: What UNC6780 Teaches AI Supply Chains
UNC6780 published malicious packages with valid SLSA attestations. Learn what provenance proves and how to build a safer AI release gate.
18 MIN READ→
Blog
Prompt Injection Is a Search Problem. Test the Agent Accordingly.
New research reframes indirect prompt injection as adaptive search. Here is a bounded method for testing tool-using agents at the action layer.
10 MIN READ→
Blog
OpenAI Has an Automated Research Intern. Secure the Experiment Loop.
OpenAI says coding agents now act like research interns. Here is a practical change-control model for agent-written code, experiments and scale decisions.
10 MIN READ→
Blog
Edge AI Attestation: Verify the Runtime Before Releasing the Model
Edge AI puts models, data and credentials on customer-owned hardware. Here is a practical attestation release gate, its limits, and the tests that matter.
10 MIN READ→
Blog
ASCII Smuggling: The Invisible Text NFKC Leaves Behind
ASCII smuggling can hide Unicode in plain-looking text. A tested Python example shows why NFKC is insufficient and what AI ingestion pipelines should check.
10 MIN READ→
Blog
Anthropic EFS Moves AI Logs Into Your Cloud. What Comes Next?
Anthropic EFS brings AI monitoring data into customer-controlled cloud storage. What CTOs should test around access, alert review, retention and recovery.
11 MIN READ→
Blog
AI Compressed a Two-Week Intrusion Into 10 Hours
Unit 42 documented an AI-assisted intrusion that crossed 50+ ATT&CK techniques in under 10 hours. Build coordinated containment across five control planes.
11 MIN READ→
Blog
OpenAI Astra Is Cyber-Critical. What Changes Now?
OpenAI has classified Astra as its first cyber-critical model. The important shift is not one benchmark score, but the access architecture required around the capability.
11 MIN READ→