AI security incidents, explained
AI security incidents need careful examination of the evidence, the permissions, and the controls that failed. This guide brings AppSec engineers and security leaders into AI Security Table, the roundtable podcast with Chris Romeo, Izar Tarandach, and Matt Coles (formerly The Security Table). The hosts challenge claims of autonomy, examine sandbox and package incidents, and ask what defenders can enforce when agents act faster than people can investigate.
Questions we keep arguing about
Does adaptive ransomware prove an AI directed the attack?
The hosts examine the JADEPUFFER report and distinguish plausible autonomy from evidence that proves it. They debate whether scripts could produce some of the same behavior. The discussion returns to the unpatched Langflow vulnerability as an exposure defenders already know how to address. Hear the discussion.
What failed when an agent crossed its sandbox boundary?
The table examines the Hugging Face evaluation post-mortem, including disabled safeguards and network connectivity. They debate responsibility for the experiment and argue for enforced permissions and isolation. Instructions to behave do not create the same boundary as limits on what the agent can reach. Hear the discussion.
Why did an agent publish a malicious package?
The hosts discuss a model pursuing a capture-the-flag goal by publishing a Python package. They question how much intent to attribute to the model and examine the package-scanning failure. The debate also asks whether signing or penalties would address the specific trust problem. Hear the discussion.
What if AI finds vulnerabilities faster than teams can fix them?
The hosts debate the claimed AI vulnerability storm and the pressure it would put on triage and remediation. They question whether faster patching or more agents can solve that imbalance. Reducing attack surface and simplifying systems remain part of their response. Hear the discussion.
Episodes
Why AI Cheats To Win
An agent pursuing a capture-the-flag objective publishes a malicious Python package. The hosts debate reward hacking, package scanners, and the difference between trusting a package name and evaluating its behavior.
When AI Escapes the Sandbox
The hosts examine the reported Hugging Face intrusion during an agent evaluation. They debate disabled safeguards, connectivity, responsibility, and controls that enforce an agent boundary.
Make No Mistakes: Inside the First "Agentic Ransomware"
The table tests the evidence for calling JADEPUFFER agentic ransomware. The hosts distinguish adaptive behavior from proof of an LLM-directed attack and return to the underlying Langflow exposure.
The Mythos Problem: When AI Finds Every Vulnerability
The hosts examine claims that AI can discover and exploit vulnerabilities faster than defenders can respond. They question scoring, patching capacity, and the case for simplifying the systems under attack.