Field notes from practitioners building and attacking AI systems in production, written for the people who have to deploy them.
Every two weeks: the AI security news, CVEs, vendor advisories, research and tooling that matter, filtered down from a few hundred raw items.
CSA Singapore's Securing Agentic AI addendum, all 114 control points, with the pieces that show each one in practice — and the gaps still waiting for an author.
A bakery's order assistant ends up writing homework and fixing code, all on the bakery's bill. Testing 199 popular custom AI apps across four no-code platforms, nine in ten could be pushed into harmful tasks and seven in ten did jobs they were never built for. How far a well-written prompt fixes that, and what a model gateway can and cannot see.
For thirty years security assumed that if something went wrong, someone had broken in. Australia's Robodebt broke none of that and still wronged hundreds of thousands of citizens. What changes once the question becomes what a system is allowed to decide.
An autonomous pentest runs in stages, and each stage can take a different model. Screening 52 models for refusal behaviour left eleven usable ones; the execution stage turned out to eat 70% of the bill. How those two findings produced three fixed profiles.
Inside Singapore's national AI Safety Red Teaming Challenge, more than 80 security experts spent three days breaking seven AI applications. What broke first, and why, says more about how AI safety is built than any single vulnerability does.
AI agents that break into systems on their own crossed a visible threshold when XBOW spent months atop HackerOne's US bug-bounty leaderboard. What "AI pentesting" actually covers, and what the benchmark scores leave out.
Published by the SGTech Cybersecurity Chapter, AI Workgroup. Pieces appear under their authors' names. If you have a practice, a failure or a lesson worth passing on, we would like to hear it.