Archive
Format: Research explained
-
Reasoning-envelope research questioned replay boundaries between model API sessions
The researchers studied whether opaque API state remained protected when reused outside its original context.
-
Anthropic’s multiagent study found conflict when agents received incompatible goals
The controlled experiments explored coordination failures, including sabotage, rather than a new breach of production Claude systems.
-
AI governance needs evidence from deployed controls, not only written policies
Qualys frames an operational evidence gap in a product article; organizations should test the claim against their own governance process.
-
An AI bill of materials is useful only when it helps answer an exposure question
Wiz’s guide describes an inventory of models, data and software dependencies, rather than an “AI Bills” vulnerability.
-
Sygnia’s 72-hour cloud intrusion investigation found familiar weaknesses at higher speed
The incident-response team assessed AI assistance from observed artifacts and activity, without identifying a novel zero-day.
-
MCP repository study shows why AI-generated vulnerability reports need verification
A large automated sweep narrowed thousands of candidates through further model checks and manual review.
-
TrendAI’s ecosystem report combines software exposure with malicious AI use
The vendor’s landscape analysis covers multiple kinds of risk; its aggregate CVE counts are not one incident or one affected product.
-
Claude Code permission research questioned fallback handling of long compound commands
Adversa’s reported finding concerned deny-rule enforcement, separate from the earlier accidental source-code release.
-
OpenAI announced plans to acquire Promptfoo for AI security testing
The March 2026 announcement concerned a security-testing business and its open-source tools, not a privilege-escalation incident.
-
AWS linked FortiGate compromises to weak credentials and exposed management ports
Amazon’s investigation found AI-assisted scaling of familiar techniques, with no FortiGate vulnerability exploitation observed in that campaign.
-
Aokigahara stealer targets AI sessions alongside developer credentials
Check Point’s reported malware analysis includes local authentication and configuration data from AI and development tools.
-
Malicious AGENTS.md file tried to turn a coding agent into a malware installer
The customer attempt was stopped before code ran; the file’s arrival route remained unknown.