Your AI Agent Is Stuck. What Is It Allowed to Do Next?
The OpenAI–Hugging Face incident, what both companies missed, and how to test your own agents. Includes a live agent poisoning scenario in Cortega Arena.
Read more →
We Ran OpenAI's Litigation Prompt
Follow-up. OpenAI published memo prompts with Astra for Law. We ran them on Llama 3.3 70B and Claude Sonnet 4.5. Bare Llama: 12 of 14 cites flagged. Bare Claude: 8 of 15 on the private-label memo, 2 of 6 on the Delaware asset-sale memo.
Read more →
A Research Plugin Is Not a Citation Check
Legal research plugins just landed in ChatGPT. That helps the model look up cases. It does not independently check the cite, or the tools the rest of the firm is already using.
Read more →
Your Chatbot Isn't Getting Jailbroken. It's Being Manipulated.
Fraud, pretexting, and policy circumvention, run live in the Cortega Arena — standard guardrails versus Cortega Policy Guardrails, same model, same prompt. Part 2 of 2.
Read more →
The Fight Between Attackers and Your Chatbot Is Completely Asymmetric
Enterprises run customer-facing AI on cheap, fast models for good economic reasons. Attackers face no such constraint. Why that cost asymmetry is the real guardrails problem — Part 1 of 2.
Read more →
FOBO Is Quietly Wrecking Your AI Model Strategy
A new “best model yet” ships every few months, and the evaluation never quite finishes. Why Fear of Better Options is driving teams to either re-benchmark on gut feel or quietly drift.
Read more →
You Will Be Educated
A “resistance is futile” school marquee turns out to be an accurate description of enterprise AI adoption. Three hard lessons on agent governance, from a billboard outside an elementary school.
Read more →
The Rockstar Prompt Is a Trojan Horse
A “5 prompts to impress your boss” download hid a prompt injection payload. Why having one LLM police another LLM’s inputs fails, and what a deterministic gateway catches instead.
Read more →
Trust Is Not a Logging Strategy
A personal AI agent misclassified a folder of financial documents, and reconstructing why turned into a real investigation — the same question a regulator asks an enterprise agent after the fact.
Read more →
Everyone Wants a Jarvis
Chief-of-Staff agent prompts are the most popular prompt search right now. Plugged into a company database or ticket queue, they are also unmonitored data conduits primed to leak enterprise data.
Read more →
Phantom FinOps: The Hidden Cost of Cheaper Tokens
Dynamic model routing can cut your AI bill 40%. It can also make debugging, compliance, and cost attribution far harder — unless you trace which model made which call, the whole way through.
Read more →
The Jared Problem: Why AI Governance Can’t Be a Policy Document
An automated resume screener learned that the two best predictors of a good hire were being named Jared and having played high-school lacrosse. Why a compliance PDF was never going to catch that.
Read more →