Hall of Pain.

Real incidents where AI agents burned real money. If your stack can do any of these, the budget gate costs less per month than 5 seconds of one of them.

DateStackLossWhat happenedSource
Jul 2025 Replit AI
(coding agent)
prod DB
+ 1,206 records
Agent issued DROP DATABASE on prod during active code freeze. Repeated instructions to not edit were ignored. CEO Amjad Masad: "unacceptable and should never be possible." Fortune
Tom's Hardware
@amasad on X
Nov 2025 LangChain + A2A
(multi-agent)
$47,000 Four agents in A2A coordination. Two got into clarification ping-pong, ran 11 straight days. $127 → $891 → $6,240 → $18,400 → $47,000 weekly. No step cap, no per-conversation budget, no orchestrator. Medium postmortem
DEV.to writeup
2026 LangGraph
(autonomous refactor)
$4,200 Solo developer kicked off an autonomous refactor over a long weekend. Three days, $4,200 in API fees, no budget circuit-breaker, workload never validated before launch. DEV.to cost-blowup walkthrough
Apr 2026 Cursor + Claude
(coding agent)
PocketOS DB
30h outage
Agent given staging credential task. Found a Railway API token in an unrelated file (not scoped). Issued one curl command — wiped production DB + all backups in 9 seconds. Rental business clients lost recent bookings, customer details, transactions. OECD.AI incident log
DevOps.com
Zenity post
⭐ burnstop is open source.   pip install burnstop  ·  github.com/santicalderon/burnstop  ·  MIT licensed  ·  CI passing  ·  drop-in proxy for the Anthropic SDK
Live runaway simulator — see the gate fire
Envelope: USD  |  Per-call: $0.0435  |  Rate: 4/sec

The math: one $47K runaway = 240 years of the Pro tier. One DB wipe = unbounded. A pre-execution budget gate at ~$0.001 per check would have caught every one of these before token-1 or call-1.

The fix: pre-execution budget gate, one line.

Rate limits cap request frequency, not budget impact. A single legitimate-looking DROP or POST call can cost more than 10K legitimate GETs. The model that loops is the one you'd ask to self-throttle — same context, same failure.

# add to your MCP client config
mcp_servers:
  cache:
    url: https://mcp.mi-kernel2026.xyz/v1/cache
    auth: $SAMURAI_KEY
    budget_envelope_usd: 5.00  # per-run hard cap
    destructive_action_requires_confirm: true

Before each tool call the cache MCP returns OK or 402 BUDGET_EXCEEDED with the reason. Your agent gets a deterministic stop signal that can't be ignored by the same model that's looping.

Pricing — built for the math above.

INDIE
$49/mo
1 project · $1K/mo prevention envelope · Slack alerts · 7-day causal replay
PRO
$199/mo
5 projects · $25K/mo envelope · 30-day replay · priority · webhooks
TEAM
$799/mo
15 seats · $250K/mo envelope · SOC2-ready audit log · SAML · 99.9% SLA
VALUE SHARE
10% saved
Add-on, capped $4,999/incident · we earn only when we prevent
Start free trial DM me your near-miss → 90d Pro free

First 20 reply-DMs from any HN agent-disaster thread get 90 days of Pro free in exchange for a 15-min call about how the incident shaped.

Built by an indie. Terms · Privacy · Agent Card