Topic

Agent security

7 articles on this topic from the NullSquare blog.

Claude

Anthropic’s Fourth Claude Incident: 7 AI Agent Containment Tests

September 10, 202610 min read

Anthropic’s Fourth Claude Incident: 7 AI Agent Containment Tests

Anthropic’s September 9 disclosure turns AI agent containment into a practical question: can you prove which actions your agent cannot take?

ClaudeAI security Agent security Security testingSandboxing

GPT-6 Astra

GPT-6 Astra and the AGI Era: Intelligence That Can Act

September 3, 20267 min read

GPT-6 Astra and the AGI Era: Intelligence That Can Act

Is GPT-6 Astra AGI? The benchmarks are mixed, but its leap in computer use, critical cyber capability, and long-horizon agency changes the security boundary.

GPT-6 AstraAGIAI security Agent securityCybersecurity

AI security

MCP security: threat model, risks, and production tests

August 28, 202614 min read

MCP security: threat model, risks, and production tests

A production-focused MCP security guide that turns tool poisoning, authorization, state, token, and runtime risks into explicit security boundaries and reproducible tests.

AI security Agent security Security testing Prompt injection

Agent security

AI agent memory needs its own security boundary

July 26, 20266 min read

AI agent memory needs its own security boundary

Persistent memory turns one hostile input into a cross-session risk. Secure the write, retrieval, action, and recovery paths as one system.

Agent security AI securityMemory poisoningPrompt injection Security testing

AI security

Persona is the hidden variable in agent behavior

July 4, 20264 min read

Persona is the hidden variable in agent behavior

NullSquare open-sourced Agent Behavior Lab so teams can measure how personas, tool definitions, histories, and model families shift agent safety.

AI security Prompt injection Agent security Model governance Security testing

AI security

Fable 5 is back. Coding agents just got harder to govern

July 4, 20266 min read

Fable 5 is back. Coding agents just got harder to govern

Anthropic restored Claude Fable 5 with tighter cyber classifiers and fallback routing. Here is what that means for coders, debuggers, and security teams.

AI security Agent security Model governance Developer ecosystems Security testing

Prompt injection

Prompt injection testing has to leave the chat box

March 21, 20261 min read

Prompt injection testing has to leave the chat box

AI security testing works best when prompt injection is tested across documents, tools, memory, permissions, and real workflow boundaries.

Prompt injection AI security Agent securityScope control