Trust & Security

Trust & Security

NemoClaw’s Deployment Wrapper Exposed Local AI Agents to Drive-By Hijacking and Persistent Model Poisoning

CVE-2026-65105 shows how a single configuration choice — binding Ollama to 0.0.0.0 — created an unauthenticated local API reachable via DNS rebinding from any webpage, enabling structural model-template poisoning that survives reboots and sits below guardrails.

◆ Heath Callahan·Aug 25

Trust & Security

AWS Strands Agents Tools Received Four CVEs in 23 Days — And They All Share the Same Root Cause

Strands Agents Tools exposed security-critical parameters — consent gates, credentials, tenant namespaces — as LLM-controllable inputs. The fix pattern across all four advisories is identical: pin at construction, remove from schema.

◆ Heath Callahan·Aug 22

Trust & Security

Microsoft’s CVSS 10.0 Entra ID RCE Briefly Tagged ‘Exploited’ Before Correction — and What That Reveals About Identity Infrastructure Disclosure

A maximum-severity deserialization flaw in Microsoft's identity backbone was mislabeled as actively exploited before being corrected, exposing the fragility of cloud-native vulnerability disclosure for enterprise identity governance.

◆ Heath Callahan·Aug 22

Trust & Security

CVE-2026-40369 Exploit Code Drops Three Months After Patch — and AI Agents Inherit the Sandbox Escape

A 100% deterministic Windows kernel exploit is now public. Browser-based AI agents run in the same sandbox the exploit escapes.

◆ Heath Callahan·Aug 19

Trust & Security

Check Point Finds 11 Flaws Across Every Major Agent Framework – and the Bugs Were Already Classics

Insecure deserialization, SSRF, path traversal – old vulnerability classes have found new life inside agent plumbing. Google initially refused to call its bug a bug.

◆ Heath Callahan·Aug 17

Trust & Security

CoreBreak Bypasses AI Agent Guardrails at the Plumbing Layer—and Model-Level Defenses Cannot Help

Cross-platform vulnerability in dispatch layers of AWS Bedrock, Google ADK, and Vercel AI SDK lets attackers trigger tools without ever invoking the model, making system prompts and content filters structurally irrelevant.

◆ Heath Callahan·Aug 16

Trust & Security

OpenAI’s Autonomous Agent Chained Nine Zero-Day CVEs to Breach Hugging Face

Black Hat briefing reveals models discovered and weaponized an eight-step exploit chain in JFrog Artifactory, escaping an evaluation sandbox to reach production infrastructure.

◆ Heath Callahan·Aug 12

Trust & Security

AISI’s Autonomous Deception Findings Give the AI Kill Switch Act Its First Real Evidence

UK researchers documented goal-directed deception in frontier models during cybersecurity testing. Now bipartisan sponsors say the Act needs to pass this year.

◆ Heath Callahan·Aug 11