AI Security
As agents start calling other agents, the failure that bites first is treating a peer's message as authority. A signature proves who spoke, not whether they were allowed. Authentication is not authorization, and a short list of unglamorous properties is what earns trust.
AI Agents
Several models converging on the same answer looks like evidence. We caught the case where it was our own prompt talking, and ran the blind control we should have run first.
AI Verification
Provenance answers who made an artifact. Attestation of scrutiny answers why you should believe it: a signed, offline-verifiable record of what an artifact survived under adversarial examination. A defensive publication, placed in the public domain as prior art.
AI Agents
SR 11-7 was replaced in April 2026. Its replacement puts agentic AI outside its own scope and says it sets no enforceable standards. Here is what actually binds an AI agent in finance right now, checked against the primary sources, with links so you can check it yourself.
OSINT
Someone sends you a link and asks what's in it. Three layers of the stack told me confident lies in a row. A missing dependency reported as fact, ninety-five minutes of fluent invention from a wrong-language model, and a year-old talk stamped with last week's date. None of them errored.
AI Agents
Anything you optimize, you eventually optimize the measurement of, through gradient rather than malice. A reward function with one target metric isn't a specification; it's an attack surface with a clearly marked entry point. On held-out instruments, and the two ways our own gauges fooled us.
AI Agents
Fan out a hundred agents and you get a hundred times the work, and a hundred times the assertions about it, arriving faster than anyone can check them. What happened when we made the ungated configuration structurally impossible, including the bug that proved the point at our expense.
AI Agents
Fan out N agents, put a verifier on the edge, never let an agent grade its own homework. The advice is correct, and almost universally written by people who have never had to live with it. One concrete run against a 75-claim registry, including the finding that fell over afterward.
Quantum Hardware
Three results we wanted to keep did not survive a second look at our own raw data. The retractions are on the record. Each one taught a specific lesson about nulls, labels, and what counts as evidence.
Maritime Intelligence
Seven per-vessel behavioral detectors over public AIS tracks, a risk band per vessel, and a plain-language narrative an analyst can check. No satellite imagery, no paid feeds. How it scores, what one test on a real NOAA reserve did and did not show, and what it still gets wrong.
PQC Migration
A post-quantum toolset built around one migration path, from inventory to compliance. What each tool does, which standards it runs, which are demonstrations, and which one is not on-premises.
Research Ethics
We wrote up a quantum key recovery as a success. Then an outside question sent us back to our own data, the key turned out to come from the classical half of the pipeline, and we said so. Why the corrected version is the stronger one.