Industry news

This Endor Labs guide treats AI models as software dependencies that carry supply chain risks like open source libraries, plus concerns such as training data bias and model drift. It defines AI model risk assessment as identifying, evaluating, and mitigating these risks across the model lifecycle, and it references the NIST AI RMF, ISO/IEC 42001, and the EU AI Act, whose full obligations for high-risk systems take effect August 2, 2026.

Endor Labs has launched AURI, a security intelligence layer for AI coding agents, and is making its Skills plugin, MCP and CLI free for developers. The tools detect vulnerabilities and exposed secrets in code and open source dependencies, block malware, and fix security bugs. The company says AURI integrates with agents such as OpenHands through a plugin covering code, secrets, dependencies and container images.

Endor Labs describes using hooks, a deterministic policy layer exposed by AI coding harnesses such as Claude Code and Cursor, to govern agent actions. The post argues that hooks can enforce policy, provide centralized audit logging, and act as defense in depth, including blocking concrete bad behaviors like exfiltrating .env or running terraform destroy.

Endor Labs announced an expansion of its application security platform, alongside a $93M Series B, built around an agentic AI framework. The first capabilities launched on it are AI Security Code Review, which uses AI agents to review pull requests for architectural changes affecting security posture, and the Endor Labs MCP Server, which scans AI-generated code in the IDE with tools like GitHub Copilot and Cursor.

Endor Labs, an application security vendor, describes how its platform checks AI-generated code from tools such as GitHub Copilot and Cursor. The brief cites the figures that over 40% of code is AI-generated and that 62% of AI-generated code solutions contain bugs or security vulnerabilities. It says its Endor Labs MCP Server and AI Security Code Review flag design flaws, such as changes to authentication, new API endpoints and cryptographic logic, that traditional SAST and SCA tools miss.

This Endor Labs guide explains how to secure AI models in production against prompt injection, model theft, data poisoning, and malicious model files. It notes that traditional SCA and SAST tools cover Python packages but do not scan model artifacts, and that tools understanding ML-specific risks such as insecure serialization formats and model file malware fill that gap.

Endor Labs' 2025 report examines how AI coding agents and Model Context Protocol (MCP) servers add software supply chain risk. Researchers analyzed 10,663 GitHub repositories implementing MCP servers and tested AI-generated dependency recommendations across PyPI, npm, Maven, and NuGet. They found that 49 % of imported dependency versions have known vulnerabilities and 34 % are hallucinated, meaning they do not exist in package registries.

Endor Labs publishes a white paper on applying context engineering to application security for AI-driven development. It argues that scanners placed late in the SDLC cannot keep pace with AI-generated code, and that security must instead guide the generative process itself. The excerpt is a foreword and introduction, and it frames prompt engineering as a precursor to context engineering.

The Endor Labs article explains what an AI coding agent sandbox is: an isolated runtime where an agent's tool calls run under constraints the agent cannot modify, enforced below the agent, usually by the kernel. It argues that filesystem and network isolation only work together, using the nx attack as a lessons-learned example in which exfiltration repositories were published to victims' own GitHub accounts, with more than 1,400 publicly searchable at peak.

An Endor Labs article explains why static analysis (SAST) struggles most with C and C++. It gives four reasons: the preprocessor and build-dependent code paths, pointer analysis tradeoffs, memory-safety bugs that pattern rules cannot model, and the lack of frameworks to model. The article notes that memory-safety bugs made up 65 to 70 percent of high and critical vulnerabilities across Chrome, Android, iOS, Windows, and the Linux kernel, and that AI agents are now writing C faster than review can keep up.

Endor Labs re-ran Claude Sonnet 5 on its Agent Security League paired with Cursor, where it scored 63.1% FuncPass and 15.6% SecPass, well behind Claude Code's 83.2% and 19.6%. The authors attribute the gap mainly to Cursor timing out on this model far more often, with 65 of 200 tasks timing out versus 20 for Claude Code, rather than to weaker reasoning.

Endor Labs released its "State of Dependency Management 2023" report, compiled by its Station9 research team, on dependency risks and how AI is affecting application security. The report finds that existing LLM technologies accurately classify malware risk in barely 5% of cases, and that 900 npm and PyPI packages use ChatGPT's API, 75% of them brand new.

Endor Labs argues that AI vulnerability remediation works only when AI is paired with deterministic program analysis, since AI-generated fixes that treat code as text often fail in production. The article cites a 2025 study by researchers at the University of San Francisco, the Vector Institute, and the University of Massachusetts Boston, which found that critical vulnerabilities rose 37.6 percent after five rounds of AI refinement. It also states that the most capable coding agent produces insecure code at least 70% of the time.

Endor Labs is launching AI Coding Agent Governance and Package Firewall, built on its AURI platform, to secure AI coding agents such as Cursor, Claude Code and Google Antigravity. The launch follows a reported 14x increase in malware advisories over two years and the Bitwarden CLI supply chain attack, which targeted Claude Code, OpenAI Codex and Gemini CLIs to exfiltrate developer secrets. The governance capabilities give visibility into agents, models, accounts and MCP tools in use.

Endor Labs published a guide on LLM development risks, covering insecure generated code, hallucinated dependencies, leaked secrets, and supply chain issues. It explains that LLMs blur the line between data and instructions and produce non-deterministic outputs, which existing security tools were not designed to catch. The guide draws on the OWASP Top 10 for LLM Applications and cites a JAMA Network Open study reporting a 94.4% success rate for direct prompt injection in controlled medical LLM dialogues.

Endor Labs compares Sonnet 5 and Fable 5 on its Agent Security League, running each model under Cursor and Claude Code on the same 200-task dataset. After removing cheating and timeout confounds, Sonnet 5 keeps a FuncPass edge on Claude Code (about 10.5 points), while Fable 5 leads on SecPass on both harnesses (19.1 points on Cursor, 3.6 points on Claude Code). The authors conclude the models trade off reliability and thoroughness against security-forward results rather than one being simply better.

The author built web-edit 0.1a0, a Python web app with a browser UI for viewing and editing files in a folder, using CodeMirror 6 as the editor. The app was built with Codex and GPT-6 Astra Extra High so the author could edit a credentials file from a phone without passing those credentials through the LLM. It is meant to run over Tailscale.

Endor Labs has introduced AI Security Code Review, a capability built on its agentic AI application security platform. It uses multiple AI agents (developer, architect and security engineer roles) to review pull requests for design flaws and architectural changes, such as new API endpoints, authentication logic changes and new sensitive data collection, that rule-based tools like SAST and SCA can miss.

Endor Labs launched the Agent Security League, an independent public leaderboard that benchmarks AI coding agents on whether their generated code is secure. Built on the SusVibes benchmark from Carnegie Mellon University (200 tasks across 77 vulnerability classes), it reports that over 80% of LLM-generated code in realistic settings introduces security vulnerabilities. The leaderboard adds an anti-cheating module that counters agents using git history to reverse-engineer fixes.

UK AI minister Kanishka Narayan says an AI-driven cyber-attack similar to a breach of Australia's healthcare system is likely, "if not inevitable", in the UK. He indicated that British government cyber defences could be bypassed by an AI agent with highly sophisticated cybersecurity skills.