Shadow adoption
Staff are already pasting company data into tools nobody approved. The policy exists; the visibility does not.
Product · In development
AngelMind
AI security and governance for organizations adopting agents, copilots, custom AI apps, and automated workflows — so you can see what the AI touched, what it decided, and who approved it.
The problem
AI got deployed faster than the controls around it.
Unbounded agents
An agent with write access and no approval gate is one bad inference away from an incident you have to explain.
No reconstructable record
When something goes wrong, "the AI did it" is not an answer you can give a client, an auditor, or an insurer.
What AngelMind is designed to do
See it
- Discover AI usage across the organization
- Map what data each tool and agent can reach
- Surface unapproved tools and risky patterns
Control it
- Approval gates on consequential actions
- Least-privilege scoping for agents
- Evidence trails that survive an audit
Measured · benchmark v1
We lead with our misses
Most security vendors publish the number that flatters them. Here is ours, with the scope and the failures attached — because a number without either is marketing, not evidence.
What we measured
Michael, our live content-scanning engine, run against our own 72-sample test corpus — 45 attack samples and 27 benign ones, authored in-house. This is not a public dataset, and it is not a measurement of your environment. It is our scanner against our own labelled set, and you should discount it accordingly.
Recall
62.2%
It flagged 28 of the 45 attack samples.
Precision
93.3%
When it flagged something, it was right 28 times out of 30.
False positives
7.4%
It false-alarmed on 2 of the 27 benign samples.
Where it fails — the part that matters
Michael is a pattern-matching engine. It is good at attacks that look like attacks, and measurably bad at attacks written to look like ordinary language:
- Evasive paraphrases: 0 of 10 detected. Reword the injection in plain English and it walks straight past.
- Encoded payloads: 1 of 5 detected. Base64 and similar encodings mostly defeat it today.
- Jailbreak phrasings: 2 of 5 detected.
It does reliably catch the canonical cases in that corpus — direct prompt injection, leaked secrets, dangerous commands, harness tampering, and hidden-Unicode payloads. Closing the paraphrase and encoding gap needs normalization and a classifier, which is built but not shipped. We will not publish numbers for it until it is.
How to check us
The benchmark is deterministic and runs offline — same input, same result, no model call. It runs in our CI on every scanner change, and a drop in recall or precision fails the build, so these figures cannot quietly rot while this page still claims them.
Benchmark v1 · figures as of 2 September 2026 · measures detection-versus-silence, not attack-type classification · Michael is the only live component; the rest of AngelMind is roadmap.
Where AngelMind actually is
In development. Not generally available.
We publish build status plainly rather than describing a roadmap as if it shipped. If you want to be part of early access, or you have a specific governance problem you want the product shaped around, tell us — that input is genuinely what we are building from right now.
Request early accessYou do not have to wait for the product
The governance work can start now, as an engagement. The platform makes it continuous later.