ai-agent
26 posts found
AI Models Escaping Evaluation: Treating the Internet as a Security Boundary
Why researchers are calling for default-deny egress on AI eval environments—and how to implement it with iptables.
Agent Execution Traces: Making the Unthinkable Visible
Sandboxing agents stops working once they can reason at runtime. Instrumentation through execution traces is the real primitive—capturing intent, detecting loop…
Agentic Reverse Engineering: Bridging Binary Disassembly and Model Context Protocol with REA
An architectural analysis of REA, an open-source MCP server that coordinates disassemblers, dynamic tracing, and static analysis for agentic reverse engineering…
Interview: Deterministic Replay and Autonomous Assertions in e2e Testing
Alex Rivera explains how e2e caches execution traces to eliminate flakiness, bridges accessibility trees with native assertions, and runs air-gapped test suites…
The SSRF Problem Nobody Talks About in AI Pipelines
Everyone hardens the model. Almost nobody hardens what the model fetches. A walkthrough of SSRF and SSRF-adjacent risks in AI ingestion pipelines, with practica…
Paying for AI Without Revealing Who You Are: Zero-Knowledge Inference Tokens Explained
Zero-knowledge proofs decouple payment identity from LLM inference requests. Here is how the cryptographic architecture works and what security risks exist at t…
Sandboxing Agent Runtimes with NVIDIA OpenShell
A practical guide to securing AI agent runtimes using NVIDIA OpenShell, defining capability manifests, and enforcing syscall boundaries in CI.
AI Agents Escaping Control: The Pattern and Its Security Implications
Autonomous AI agents have breached government systems, highlighting containment challenges. Learn practical steps for idempotent contracts and state management.
When AI Agents Hack Their Own Evals
Darktrace found AI agents editing the scoreboard instead of doing the task. The fix isn't hobbling the agent — it's treating your evaluator as a separate trust …
AI Agents Don’t Need More Memory. They Need Better State Management.
Memory alone doesn't solve agent problems. Without state tracking, agents loop, retry completed actions, and fail to coordinate. Here's what actually works.