Alice in Supply Chains podcast

Episode 22 | October, 2026

0:00
52:53
Rewind 15 seconds
Fast Forward 15 seconds

This month, Alexandre Sieira and Adrian Sanabria open with a story that punctures some of the AI hype: most of the recent headlines about AI agents "escaping" and hacking other organizations trace back not to superintelligent models, but to a single misconfiguration at Irregular, the third-party firm Google, Meta, Anthropic, and OpenAI used to safely test their models. By Adrian's tally, four of the seven known incidents stem from that one mistake, and only OpenAI's were genuine sandbox escapes. It's a textbook TPCRM lesson: shared responsibility doesn't mean your vendor is getting their half right.Story two is the Manchester Airports Group breach, where attackers abused a read-write API key for Iterable, a marketing tool used just to track email clicks, that sat in the site's client-side JavaScript for four years despite Iterable's documentation explicitly saying never to expose it. Alexandre uses the case to highlight the "hidden third parties" holding your executives' travel data (airports, airlines, hotels) that never pass through your TPCRM process.The episode closes with legal AI company Harvey's shift from frontier models to cheaper open-weight models, which Alexandre uses to untangle two decisions that often get conflated: which model you use, and where its inference runs. Each carries its own jurisdictional, security, and provenance risks, from where your data gets processed to whether a model's training data was poisoned (research suggests as few as 250 documents may be enough). The pair also dig into TypeSafe's new Jev model, built for calibrated decisions rather than text generation. It's hundreds of times faster and cheaper for classification tasks like spotting malicious traffic or routing prompts between models, but its "hallucination-free" claim doesn't hold up: a confidently wrong answer is still wrong, and harder to spot.Links:-Story 1: https://cyberscoop.com/irregular-ai-s...-Story 2: https://scotthelme.co.uk/no-hacking-r...-Story 3: https://www.bloomberg.com/news/articl...-The annual Tenchi conference is coming up soon: https://www.tenchisecurity.com/en/com...-The creator of Jev presents his creation at an AI conference and explains why LLMs aren’t good at automation    • Jev CEO: I made ChatGPT, now I'm building ...  -Alexandre’s LinkedIn post with his prediction about Jev https://lnkd.in/p/dAXiGdye

More episodes from "Alice in Supply Chains"