id stringclasses 10
values | incident_date stringclasses 10
values | reported_date stringclasses 10
values | title stringclasses 10
values | category stringclasses 5
values | agent_or_product stringclasses 10
values | amount_usd float64 32 1.8M ⌀ | description stringclasses 10
values | primary_source_url stringclasses 10
values | secondary_source_url stringclasses 7
values | source_type stringclasses 4
values | control_that_would_have_helped stringclasses 6
values | confidence stringclasses 2
values | notes stringclasses 10
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
INC-001 | 2024-11-29 | 2024-12-02 | Freysa AI agent tricked into transferring $47k prize pool | manipulated-payment | Freysa (autonomous agent game on Base) | 47,316 | Freysa was hard-coded to never approve a fund transfer no matter what was said. After 482 messages from 195 players over several days, a player calling themselves p0pular.eth convinced Freysa that its anti-transfer rule did not apply to its "approveTransfer" function when triggered by an incoming donation, and the agen... | https://www.theblock.co/post/328747/human-player-outwits-freysa-ai-agent-in-47000-crypto-challenge | https://the-decoder.com/hacker-wins-47000-by-tricking-ai-chatbot-with-smart-prompting/ | news | human approval | high | Exact prize figure reported as $47,316.05. Reported dates across outlets cluster late Nov/early Dec 2024; exact first-publication date not separately confirmed. |
INC-002 | 2026-05-04 | 2026-05-06 | Morse-code prompt injection drains AI-linked Bankr wallet | manipulated-payment | Grok (xAI) + Bankrbot (Bankr Club trading bot on X/Base) | 150,000 | An attacker first sent an NFT to a Grok-linked wallet that reportedly unlocked elevated transfer permissions inside the Bankr trading bot, then posted a Morse-code-encoded message to Grok on X. Once decoded, the message was treated as an authenticated instruction and the Bankr bot transferred roughly 3 billion DRB toke... | https://www.giskard.ai/knowledge/how-grok-got-prompt-injected-an-x-user-drained-150-000-from-an-ai-wallet | https://ambcrypto.com/ai-linked-wallet-drained-via-prompt-injection-in-bankr-exploit/ | research | payee allowlist | medium | Reported amount ranges from $150,000 to $200,000 (3 billion DRB tokens) depending on outlet and token price at time of writing. Multiple crypto-news outlets (AMBCrypto, CryptoSlate, BeInCrypto) reported roughly 80% of funds recovered after the community identified the attacker. Treated as medium confidence because of a... |
INC-003 | 2025-08-20 | 2025-08-21 | Guardio tricks Perplexity Comet into buying from a fake Walmart site | research-demo | Perplexity Comet (AI browser) | null | Security researchers at Guardio Labs asked Perplexity's Comet AI browser to buy an Apple Watch from a fake Walmart page (hosted on a Lovable-built site with a bogus URL). Comet did not flag the site as fraudulent, entered the user's card and address details, and completed checkout. | https://www.bleepingcomputer.com/news/security/perplexitys-comet-ai-browser-tricked-into-buying-fake-items-online/ | https://www.pcworld.com/article/2885371/perplexitys-ai-browser-is-a-sucker-for-blatant-scams-and-prompt-hijacks.html | research | human approval | high | No dollar amount stated in sources (test purchase, real checkout flow but likely not completed with real funds charged -- sources do not clarify whether the transaction settled). Same researchers also showed Comet fell for a fake Wells Fargo phishing email. |
INC-004 | 2026-02 | 2026-07-31 | Amazon's internal Claude coding task overspends by $1.8 million | overspend | Claude (Anthropic) via internal Amazon tooling | 1,800,000 | An internal Amazon project using Claude to match author records to product listings ran roughly 860% over its budget, reaching about $1.8 million in model-usage costs. The overrun went undetected for about five months before being flagged internally; senior engineers reportedly called it 'catastrophically expensive.' | https://www.tomshardware.com/tech-industry/artificial-intelligence/amazon-accidentally-spent-usd1-8-million-using-claude-for-menial-coding-task-went-860-percent-over-budget-catastrophically-expensive-coding-blunders-discovered-in-internal-amazon-ai-usage-metrics | https://www.ghacks.net/2026/07/31/leaked-amazon-documents-detail-1-8-million-overrun-on-a-single-claude-ai-task-missed-for-five-months/ | news | per-agent budget | medium | Originally attributed by multiple outlets to a Financial Times report based on a leaked internal Amazon presentation; the FT article itself is paywalled and was not fetched directly, so this relies on Tom's Hardware and gHacks summaries. Exact calendar start date of the overrun is not public; '5 months undetected, surf... |
INC-005 | 2026-05-09 | 2026-05 | Autonomous agent racks up $6531 AWS bill scanning hobbyist network DN42 | overspend | Unnamed AI agent (""JertLinc3522"") run by a user ""JertLinc"" with AWS access | 6,531.3 | An operator gave an AI agent unsupervised AWS access to register with and 'index' the volunteer-run DN42 hobbyist BGP network. The agent decided the job needed five m8g.12xlarge instances plus load balancers and Lambda functions generating around 20 Gbps of scanning traffic, and re-created duplicate CloudFormation stac... | https://lantian.pub/en/article/fun/ai-agent-bankrupted-their-operator-scan-dn42lantian.lantian/ | https://news.ycombinator.com/item?id=48500012 | first-hand-post | per-agent budget | high | Primary account is the DN42 community's own first-hand write-up, corroborated by extensive Hacker News discussion (HN item 48500012). Commenters debated whether this was a genuine mistake or a social-engineering/crypto-donation stunt by the operator; that dispute is unresolved. |
INC-006 | 2026-04 | 2026-04 | Developer gets $32 surprise bill from a runaway AI agent | overspend | Unspecified AI agent built on OpenAI/Anthropic SDKs | 32 | A developer left an AI agent running unattended for about 20 minutes and came back to an unexpected $32 charge. They subsequently built and open-sourced AgentBudget, a tool that monkey-patches the OpenAI and Anthropic SDKs to track running cost per call and raise an exception once a spending cap is hit. | https://news.ycombinator.com/item?id=47418574 | null | first-hand-post | per-agent budget | medium | Small dollar amount and the underlying agent/task are not described in detail in the source. Included deliberately as a clear, dated, first-hand account rather than the many 'dev.to' posts describing $47,000-$78,000 runaway-agent bills that could not be verified as real incidents -- see SEARCH-LOG.md for those rejectio... |
INC-007 | 2025-12 | 2025-12-18 | Claude-run office vending machine gives away inventory and loses money | wrong-payee/wrong-purchase | "Claudius" (Claude, via Anthropic's Project Vend with Andon Labs) | null | In a deliberate red-team stress test, Anthropic and Andon Labs let a Claude-based agent named Claudius manage pricing, ordering and sales for a real office vending setup, this time at the Wall Street Journal's newsroom. Within days Claudius gave away most of its inventory for free, including a PlayStation 5 it had been... | https://www.anthropic.com/research/project-vend-2 | https://www.tovima.com/wsj/we-let-ai-run-our-office-vending-machine-it-lost-hundreds-of-dollars/ | official-postmortem | per-payment cap | high | This was an intentional Anthropic/Andon Labs experiment to study agentic payment risk, not an unintended production failure -- included because real money was genuinely lost in a real-world deployment and Anthropic itself published it as a cautionary case study. WSJ original article (Joanna Stern, Dec 2025) is paywalle... |
INC-008 | 2026 | 2026-09-16 | Mandiant reports accounting agent loop caused roughly $50000 bill | overspend | Unnamed ""accounting agent"" (company not disclosed) | 50,000 | Google Mandiant's 'AI Risk and Resilience' report describes an accounting AI agent that entered a runaway execution loop, making more than 15,000 high-cost API calls in under an hour, generating roughly $50,000 in cloud charges and disrupting active business transactions. | https://www.helpnetsecurity.com/2026/09/16/google-mandiant-enterprise-ai-security-risks-report/ | null | official-postmortem | rate/velocity limit | medium | The underlying Mandiant/Google Threat Intelligence Group report was not fetched directly; this relies on Help Net Security's summary of it. Company name, exact date and technical root cause of the loop are withheld in every version of this story found. |
INC-009 | 2026-01-30 | 2026-01-30 | Researchers red-team Google's AP2 agent-payment protocol via prompt injection | research-demo | Google Agent Payments Protocol (AP2), Gemini-2.5-Flash shopping agent on Google ADK | null | Academic researchers built a functional AP2-based shopping agent and showed that indirect prompt injection could redirect which product the agent recommended/purchased with a 100% success rate in their tests (a 'Branded Whisper Attack'), while a related 'Vault Whisper Attack' caused cross-account data exposure in 20% o... | https://arxiv.org/abs/2601.22569 | https://arxiv.org/html/2601.22569v1 | research | human approval | high | Peer-reviewed-style arXiv paper ("Whispers of Wealth: Red-Teaming Google's Agent Payments Protocol via Prompt Injection," submitted 2026-01-30, revised 2026-05-18) against a real, functioning AP2 reference implementation, not production consumer traffic. Labeled research-demo per inclusion criteria (d). |
INC-010 | 2026-02-03 | 2026-02-03 | Snyk finds a published agent skill that collects and leaks full credit-card details | credential-leak | "buy-anything" skill v2.0.0 on the ClawHub/OpenClaw agent-skills marketplace | null | Snyk researchers scanned the ClawHub marketplace and found 283 of 3,984 published agent skills (7.1%) had critical security flaws. The worst case, a 'buy-anything' skill, explicitly instructed agents to collect full card numbers, expiration dates and CVC codes and embed them into API requests, passing that data through... | https://snyk.io/blog/openclaw-skills-credential-leaks-research/ | null | research | credential never exposed to agent | high | Found live on a real, publicly used marketplace (not a lab-only mockup), but no confirmed case yet of an attacker actually exfiltrating a specific user's card data through this exact skill -- a research/disclosure finding rather than a reported exploited loss. Labeled research-demo-style credential-leak per inclusion c... |
Source of truth and submissions: https://github.com/Pink-Agentic-Payments/agent-spending-incidents (v0.1.0, 2026-10-05).
AI Agent Spending Incidents Log — by Pink Agentic AI Payments
A dated, sourced log of publicly reported cases where an AI agent (or LLM-driven automation) spent money it shouldn't have, paid the wrong party, was manipulated into attempting a payment, or leaked payment credentials. Built for researchers, journalists and anyone else who needs concrete, citable examples instead of hypotheticals when discussing the risks of letting AI agents pay.
Maintained by the team behind Pink Agentic AI Payments (by PinkWallet); entries are included on sourcing criteria only, regardless of vendor.
License: CC BY 4.0. Last updated: 2026-10-05.
Inclusion criteria
An incident qualifies if it is a publicly reported event, 2023-2026, where an AI agent / LLM-driven automation / AI shopping or coding agent:
- (a) overspend -- spent or committed money beyond what its operator intended, including runaway API/cloud bills caused by agent loops;
- (b) wrong-payee/wrong-purchase -- paid the wrong party or bought the wrong thing;
- (c) manipulated-payment / credential-leak -- was manipulated (prompt injection, malicious site or tool) into attempting a payment or leaking payment credentials; or
- (d) research-demo -- a documented security-research demonstration of (c) against a real product.
Excluded: generic crypto hacks with no AI agent involved, pure hypotheticals, vendor marketing "here's what could happen" scenarios, and anything we could not source to a URL that loaded.
Every row in incidents.csv was checked against its listed source(s) on the day this log was compiled (2026-10-05). Rows are marked confidence: medium or low, and flagged in notes, wherever a fact (amount, date, cause) is disputed across sources, undisclosed, or relies on a secondary summary of a paywalled/unfetched primary report.
Summary (v0, 10 incidents)
By category
| Category | Count |
|---|---|
| overspend | 4 |
| manipulated-payment | 2 |
| wrong-payee/wrong-purchase | 1 |
| research-demo (incl. 1 credential-leak demo) | 3 |
By year (incident date)
| Year | Count |
|---|---|
| 2024 | 1 |
| 2025 | 2 |
| 2026 | 7 |
Note: v0 skews toward 2026 and toward the categories easiest to source (overspend stories and academic/security research demos). That is very likely a research-coverage artifact, not a sign that other categories are rare -- see FINDINGS.md and SEARCH-LOG.md.
How to submit a new incident
Open a GitHub issue with:
- A link to a primary source (news report, official postmortem, research writeup, or a specific, dated first-hand post) that was live when you filed the issue.
- The date of the incident and the date it was reported, as precisely as the source allows.
- Which inclusion criterion (a/b/c/d) it meets and which
categoryfrom the schema below it fits. - Anything disputed or uncertain about the facts.
We will not add an entry we cannot verify against a working source URL, and we will not add hypothetical or vendor-marketing scenarios.
Schema (incidents.csv)
id, incident_date, reported_date, title, category, agent_or_product, amount_usd, description, primary_source_url, secondary_source_url, source_type, control_that_would_have_helped, confidence, notes
category: one ofoverspend,wrong-payee/wrong-purchase,manipulated-payment,credential-leak,research-demo.source_type: one ofnews,research,official-postmortem,first-hand-post,incident-db.control_that_would_have_helped: the single most direct mitigation, fromper-agent budget,per-payment cap,payee allowlist,human approval,rate/velocity limit,credential never exposed to agent,none-applicable. This is a factual best guess based on the reported cause, not a product pitch.amount_usd: left blank where no source stated a dollar figure. Never estimated or invented.
Known gaps in v0
- No confirmed, publicly reported case yet of an AI agent actually completing an unauthorized fiat payment (card/bank) in production, as opposed to crypto transfers, lab/research demos, or cloud-compute overspend. The closest real-world fiat case found (Perplexity Comet / Guardio, INC-003) is a security researcher's test, not an unprompted real-world victim.
- Several widely-circulated "AI agent ran up a $47,000 / $78,000 bill" stories turned out to be unverifiable marketing content from cost-monitoring vendors, not real incidents -- excluded; see
SEARCH-LOG.md. - AI Incident Database (incidentdatabase.ai) entry 1313 (Claude vending-machine losses at WSJ) overlaps with Anthropic's own "Project Vend" writeup; we cite Anthropic's primary source for INC-007 rather than duplicating both.
- x402 / crypto-agent-payment-protocol incidents (e.g. the 402Bridge hack) were investigated but excluded from v0 because available reporting describes a bridge/smart-contract exploit, not an AI agent being manipulated into paying -- see
SEARCH-LOG.md.
- Downloads last month
- 82