NullSquare blog

Field notes for continuous offensive security.

Practical writing on AI-assisted assessments, release gates, attack surface coverage, evidence, and the operating habits that keep security testing useful after the report lands.

AI security14 Agent security7 Security testing7 Continuous testing5 Prompt injection5 Evidence4 Model governance4 Attack surface3 Retesting3 Developer ecosystems2

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?

Read article

Latest articles

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

AI security

Automated Penetration Testing: A Practical Buyer’s Guide

August 18, 20268 min read

Automated Penetration Testing: A Practical Buyer’s Guide

How to evaluate automated and AI penetration testing platforms by scope control, safety, evidence, coverage, retesting, and human oversight.

AI security Continuous testing Evidence Retesting

AI security

AI vulnerability remediation: what Chrome’s 1,072 fixes reveal

July 31, 20265 min read

AI vulnerability remediation: what Chrome’s 1,072 fixes reveal

Google’s Chrome security update shows why AI vulnerability remediation must measure validated fixes, deployment coverage, and exploit retesting.

AI securityAI remediationVulnerability managementContinuous testingCode security

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

Compliance

Compliance tools are not all solving the same problem

July 15, 20262 min read

Compliance tools are not all solving the same problem

A practical comparison of compliance automation platforms, where NullSquare fits, and why evidence review depth matters as much as audit workflow coverage.

ComplianceGRCEvidence AI securitySecurity operations

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

AI security

GPT-6 restrictions could cost the U.S. the AI race

June 29, 20267 min read

GPT-6 restrictions could cost the U.S. the AI race

A data-backed argument for why blanket GPT-6 access restrictions would weaken U.S. AI leadership, and what targeted release gates should replace them.

AI securityAI policyModel governance Security testing Developer ecosystems

AI safety

The Fable ban is really a scope-control warning

June 20, 20263 min read

The Fable ban is really a scope-control warning

Anthropic Fable 5 showed the hard truth of frontier AI safety: stronger coding and bug-finding models are also stronger cyber systems.

AI safetyPrompt injectionCyber capabilityModel governance Security testing

Evidence

Evidence quality is the real security signal

May 29, 20261 min read

Evidence quality is the real security signal

Security findings only create momentum when evidence is reproducible, scoped, understandable, and easy to retest after a fix lands.

EvidenceFindingsRetesting Continuous testing

Private runners

Private runners make internal testing practical

May 8, 20261 min read

Private runners make internal testing practical

Internal security testing needs local reach, strict authorization, and evidence handling that works without exposing private networks.

Private runnersInternal testingAttack surface Continuous testing

AI security

The annual pentest myth breaks in the glass-door era

April 18, 20261 min read

The annual pentest myth breaks in the glass-door era

When AI agents can attack and defend continuously, companies need living security evidence instead of one annual pentest report.

AI securityContinuous monitoringPentestingAttack surfaceDefensive AI

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

AI security

AI security testing works best when it becomes a release gate

February 27, 20261 min read

AI security testing works best when it becomes a release gate

A practical model for using AI-assisted security testing before release without slowing engineering teams down.

AI securityRelease gatesSDLC

AI security

Continuous AI security testing: how the loop actually works

January 15, 20265 min read

Continuous AI security testing: how the loop actually works

Continuous AI security testing is a loop, not a faster scan. Here is what each stage has to do, and how a finding earns the right to close.

AI security Continuous testing Attack surface Evidence Retesting