BLACK_WALL failure modesbenchmarkhome
AGENT STRESS TEST

Stress-test your agent before it ships.

You wouldn't ship code without tests. But your AI agent doesn't just return text — it sends, pays, runs SQL, deletes. Run the actions it can take through 28 documented failure modes and find what it'd wave through before a customer does.

28
FAILURE MODES TESTED
0
INSTALL — RUNS PRE-DEPLOY
GO·STOP
VERDICT + THE FIX PER ACTION

Why pre-flight an agent

Every agent disaster this year was an action nothing checked first. PocketOS: a clean agent deleted prod + every backup in 9 seconds. The code wasn't buggy — it executed. A stress test surfaces those before launch: you point Black_Wall at the actions your agent can take, and it tells you which ones it would let through that it shouldn't.

How it works

  1. List the risky actions your agent can perform — send_email, run_sql, make_payment, delete_file, post_content, api_call.
  2. Black_Wall forecasts each against the 28 failure modes — predicted outcome, blast radius, red flags, and a GO / CONFIRM / STOP verdict.
  3. You get a report card — which actions your agent would wave through that it shouldn't, why, and the safer alternative for each.

Example report card

research-crew-demo · 6 actions checked3 would ship unchecked
STOPrun_sql · DROP TABLE — DESTRUCTIVE_VERB · gated (would've executed unchecked)
STOPdelete_file · /prod — IRREVERSIBLE_NO_BACKUP · gated
CONFIRMmake_payment · $48k — AMOUNT_OUT_OF_BAND · held for a human
GOsend_email · teammate recap — cleared to proceed
GOrun_sql · SELECT count(*) — cleared
GOread_file · /var/log — cleared

Illustrative example. Your report runs against your agent's actions. The gate's own results on a labeled set are public on the benchmark.

Run it three ways

  1. One action, right now — paste it into the live demo, no signup, instant verdict.
  2. The full battery — run your agent's action set through the API or MCP server with a free key (~100 forecasts/mo, no card).
  3. Live near-misses — drop Black_Wall in observe mode for a week (it logs, never blocks) and get a report of every destructive thing your agent almost did in production.

Stress-test your first action in 10 seconds — no signup.

Try it live →Get a free key