BLACKBOX: a flight recorder for AI agents

BLACKBOX is a flight recorder for AI agents. Every step an agent takes is appended to a hash-chained log: each record's hash covers the record and the hash before it, so the log cannot be edited, dropped, inserted or reordered afterwards without every later hash going wrong. The same log renders to one self-contained HTML file: a 4D flight path (tool × time × run) you can scrub, a timeline, and an in-browser re-verification of the whole chain with WebCrypto. Live: greenaidigital.com/blackbox.

It ships three ways: a pure standard-library Python recorder (wrap any callable, secrets scrubbed by key name and by shape), an MCP server with five tools so any agent host can record its own runs, and a BuildShip node in JavaScript with byte-identical hashes, proven by a test in the repo. Source, MIT: github.com/GreenAiSolution/blackbox.

BULWARK, my provenance-labelled agent, ran ten jobs × three prompt-injection attacks × five delivery channels: 150 runs, 602 records.

0 attacks landed, 149 of 150 jobs done, 1 run refused by policy (the poisoned invoice). Every tool call, its labels and the refusal reason are in the log.

Verified: OK 602 records, sealed. Then I edited record 40 on disk: BROKEN at record 40: content edited after it was written. The replay page has a tamper button that does the same thing live.

The Flight-Recorded Agent is my flagship: I build your agent (your stack or BuildShip), run my injection bench against it across six channels, and ship it with BLACKBOX strapped on from day one, so your customers can replay and re-verify any run. Fixed scope, three weeks, you own every line. Listed on my profile as an enhanced service.

Not secrecy: the log is plaintext by design; secrets are scrubbed before writing.

Not a signature: anyone can write a valid chain. The guarantee is that nobody can change one afterwards without it showing. A signed head hash proves authorship; that is a one-line addition.

Not a tracer: it records what you wrap. Wrap the tools and the policy checks and it records everything that matters.

Related: the attack lattice it grew out of, Tesseract, and the lab where you play the attacker, The Gauntlet.