7 articles tagged
OpenAI's internal red-teamer beat human red-teamers 84% to 13% and broke a live vending-machine agent. GPT-5.6 Sol now fails 0.05% of those attacks. Here is why a hardened model is not a defence plan, and the seven
GitHub shipped a CodeQL query on 10 July 2026 that catches prompt injection in code review: js/system-prompt-injection flags untrusted input flowing into an AI model's system prompt. Here is what it detects, how to turn
Two Cursor flaws rated 9.8 let a single prompt escape the editor's sandbox and run commands as you. Capsem takes a different line: a hardware-isolated Linux VM per agent session, with every HTTPS request inspected
Prompt injection in agentic browsers like Atlas and Comet has no reliable fix, and Atlas stops working on 9 August 2026. Here is a rollout decision and 7 controls, from isolation to DLP, for security teams.
Adversa AI tested 11 open-source coding agents in June 2026 and beat the safety filter on 10 of them with shell tricks that are 30 years old. How the bypass works, which designs held, what to change now.
Prompt injection is still OWASP's #1 LLM vulnerability in 2026, and agentic AI turns it into fraud and data theft. This guide shows how to contain it with least privilege, human approval and sandboxed tools.
Most AI agent projects fail on value, security and cost, not on the model. Here are five engineering lessons CTOs learn the hard way in 2026, backed by Gartner, MIT and OWASP data, from scoping ROI to controlling token
