Manus Ran Email Code Before Warning the User in Salt’s Test
Salt's Manus test: code ran before warning; flaw resolved.
AI
Artificial intelligence, local AI, agents, tooling and AI security.
GitLab’s Self-Hosted AI Gateway Needs an Immediate Update
Self-hosted gateway fix available.
Google’s AI Study Supports Exposure-Based Patching
Prioritise AI service fixes by verified deployment exposure.
Google Opens Gemini 4 Argon to Selected Cyber Defenders
Give findings and patches an accountable review path.
Two LightLLM Helper Ports Expose Code Execution on Inference Nodes
Two LightLLM helper services can accept unsafe network input outside the model API. Check multimodal and profiling nodes, their live ports and their network boundaries.
A Patched MCP Client Can Still Send Its Secret to the Wrong Server
A fixed MCP Python SDK release is only part of the repair for unattended OAuth clients. Pin the issuer, rebind saved registrations and assess possible credential exposure.
The AI Agent Copied the Attack Into Its Own Reply
In an OpenAI test, a routine email reply carried a prompt injection into the outgoing message. The result was simulated, but the path between agents deserves a real-world test.
Your Salesforce Agent Could Have Sent the Phish in Slack
A malicious CRM lead could cross into Agentforce and emerge as a trusted Slack reply without showing who triggered it. Salesforce has changed the defaults.
The Hidden ChatGPT Task That Read Gmail While the User Saw a Normal Answer
Check Point used a shared internal package service to pass hidden tasks between separate ChatGPT accounts. In its demonstration, the victim saw a normal answer while Gmail data crossed the boundary.
AI Kept Rebuilding This Russia-Linked Malware Until Security Tools Missed It
Anthropic says a Russia-linked espionage operator used AI to monitor detections and keep rebuilding malware until it was no longer detected. The attack cycle, not a single hash, is now the defensive problem.