building Air-Blackbox
San Diego, CA
Joined March 2026
- Tweets94
- Following93
- Followers11
- Likes6
Most AI recruiting tools keep no record of what the agent actually did.
Mine does now.
This is my sourcing product writing a tamper-evident receipt for every decision it makes about a person. Real search, real run, nothing mocked:
UK safety evaluators watched AI agents take 19 unsanctioned actions including inventing fake people to trick real ones.
"Unsanctioned" is doing a lot of work in that sentence.
Sanctioned by what?
Here's what a machine-readable sanction looks like and a cover-up getting caught:
Honest limit: this doesn't stop a model from saying something false.
It stops it from doing unsanctioned things at the tool boundary, and proves what happened either way.
You can still lie. You just can't do it invisibly.
Shipped v1.14.0 this week. Open source, Apache 2.0.
Don't take my word for it — it ships an attack harness that runs 5 real attacks and prints the scorecard:
pip install air-blackbox
python demo/hiring_audit_day.py
github.com/airblackbox/airbl…
The demo in the video is real code, not a mockup:
- agent asked to guess ethnicity from a name → BLOCKED before execution, attempt recorded
- 6 months later, history rewritten + re-signed → internal check passes
- the outside witness catches it anyway
AI agents loop, run away, and delete things they shouldn't. I got tired of watching it happen, so I built a guard for it.
Tombstone: stop your agent from looping, running away, or wiping data. In two lines. With a receipt.
github.com/airblackbox/tombs…
You gave an AI agent access to your file system. Right now you're trusting it not to delete the wrong thing.
Observability tools show you the disaster in a dashboard. After it happens.Tombstone stops its hand before it lands, then hands you a signed receipt of what it tried.
An agent blocked from deleting real files. A runaway loop killed at step 5. A receipt that verifies. Runs locally, your data never leaves.
Agents get more autonomous every week. The question isn't whether one fires a destructive action. It's whether you stopped it, and whether you can prove it.
Open source. Apache 2.0. github.com/airblackbox/tombs…
I built a tamper-evident audit chain for AI decisions. So I put it in your browser.
Tamper with one record. Watch the cryptographic proof collapse, live.
Not an animation. Real HMAC-SHA256 + Ed25519, the same chain `pip install air-blackbox` writes.
airblackbox.ai/demo/ledger.h…