Non-Repudiation Needs More Than a Signature
A signed action can prove origin and integrity without proving informed authority or a fair decision process.
AgentCollusionAI security startupNews
News.
Product updates, research, and original films. Follow what we build and what we learn about systems of agents.
Human-Present and Human-Absent Authority Differ
A user approving a final action creates different evidence from a standing intent executed later by agents.
MCP Token Passthrough Is Dangerous
Forwarding the client’s bearer token to an upstream API erases audience boundaries and creates confused deputies.
Resource Indicators Bound Agent Credentials
OAuth resource parameters help ensure a token issued for one tool cannot be replayed at another.
Tool Descriptions Are Part of the Attack Surface
Agents often choose tools from natural-language metadata that can redirect plans before any protected call occurs.
MCP Discovery Needs Issuer Validation
Discovering an authorization server is useful only if the client can prove it belongs to the intended protected resource.
MCP Deprecations Change Security Assumptions
When protocol features move or disappear, monitoring and authorization designs built around them can silently weaken.
Stateless MCP Moves State Into the Evidence Layer
Removing transport sessions improves scalability but forces systems to preserve task and principal continuity elsewhere.
Task Cancellation Is a Security Control
Cancellation is not merely user experience; it is the mechanism that limits an autonomous plan after risk changes.
Tool Outputs Need Provenance Before Reuse
An agent may treat a tool result as fact even after it has crossed servers, summaries, memories, and other agents.
