Audit log

Every state-changing event for AI Hallucinations: moderation decisions on community submissions, plus corrections and updates from the news pipeline. URL-based decisions carry three independent witnesses — the original source, an Internet Archive snapshot taken at submission time, and a Solana memo signed by our publicly-disclosed publisher key.

1. #1publishby system:backfill

   2026-05-28 17:26:04Z
   Score: ? → ? (no score change)
   anchoranchored

   chain

   ●mainnet-betaslot 422,765,306

   sig

   MgJJWcHoiyzq…u7ttXKUTcopyexplorer ↗

   hash

   7M1mTmRTwWzC…xZ497ZXKcopysha256 → base58

   verifying row…full verify ↗

   canonical bytes (28720 B) ▸

   {"actor":"system:backfill","investigation_id":"a8098fd1-f363-46c0-9c75-1bf965e72753","kind":"publish","page_slug":"hallucinations","published_at":"2026-05-28T17:26:04.143Z","sequence_num":1,"snapshot":{"content_type":"investigation","entity_name":"AI Hallucinations","sections":[{"content":"An AI hallucination occurs when a large language model (LLM) generates information that is inaccurate or entirely fabricated, yet is presented with the same linguistic confidence as verified fact. Unlike human errors, which often carry hedging language or uncertainty markers, LLM hallucinations typically appear fluent, specific, and authoritative. Zealynx's blockchain security glossary identifies four primary types relevant to crypto contexts: factual hallucinations (false claims such as fabricated CVE numbers or non-existent exploits), logical hallucinations (internally inconsistent reasoning), attribution hallucinations (fake citations, fabricated research papers, non-existent URLs), and temporal hallucinations (misrepresented timelines critical to governance and security incidents). The root technical causes include models lacking grounding in external real-time knowledge, reinforcement learning incentives that reward confident and detailed answers over cautious accurate ones, and sampling randomness that introduces errors during token generation. A September 2025 paper published on arXiv classified hallucinations not as an incidental defect but as a structural property of current LLM architectures, stemming directly from their statistical estimation process.","heading":"Definition and Technical Background","severity":"high","sources":[{"credibility":2,"name":"AI Hallucination — Blockchain Security Glossary, Zealynx","type":"research","url":"https://www.zealynx.io/glossary/ai-hallucination"},{"credibility":2,"name":"Agentic AI and Hallucinations (arXiv 2507.19183)","type":"research","url":"https://arxiv.org/pdf/2507.19183"}]},{"content":"A peer-reviewed benchmark published on arXiv in October 2025 — 'When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets' (arXiv 2510.00332) — evaluated 17 AI models across 178 time-anchored tasks using cryptocurrency markets as the test environment, citing the $30 billion lost to exploits in 2024 as the stakes context. Without tool access, frontier models achieved only 28% accuracy on tasks described as routine for junior analysts. With full tool augmentation, performance plateaued at 67.4% against an 80% human baseline. The study identified a critical failure mode: models systematically preferred unreliable web search over authoritative data sources, repeatedly falling for what the authors described as SEO-optimized misinformation and social media manipulation. Notably, not a single model across all 17 evaluated successfully completed a baseline task involving blockchain data retrieval — despite having access to specialized tools explicitly documented for that purpose. A separate December 2024 arXiv paper (2512.03107) on information-theoretic hallucination detection in finance found that even with mitigation, standard LLMs exhibited substantial hallucination rates on financial question-answering tasks, and proposed the ECLIPSE framework which achieved a 92% relative reduction in hallucination rate at 30% coverage.","heading":"Academic Benchmarking: Performance Gaps in Crypto Contexts","severity":"critical","sources":[{"credibility":2,"name":"When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets (arXiv 2510.00332)","type":"research","url":"https://arxiv.org/abs/2510.00332"},{"credibility":2,"name":"Detecting AI Hallucinations in Finance: An Information-Theoretic Method Cuts Hallucination Rate by 92% (arXiv 2512.03107)","type":"research","url":"https://arxiv.org/abs/2512.03107"}]},{"content":"Independent benchmarking by Vectara using the HHEM 2.1 evaluation framework found DeepSeek-R1 to exhibit a 14.3% hallucination rate — approximately four times higher than its predecessor DeepSeek-V3, which scored 3.9%. Vectara attributed this to R1's tendency to overhelp by introducing unsupported facts into responses beyond the source material. This is directly relevant to the cryptocurrency sector because hundreds of AI agent tokens — autonomous on-chain agents whose market capitalizations reached above $70 billion in early 2025 — rely on reasoning-style LLMs for trading signals, protocol research, and smart contract execution. Tokens including Virtuals Protocol (VIRTUAL), ai16z (AI16Z), and aixbt (AIXBT) are among the projects exposed to underlying model reliability gaps. BeInCrypto reported in early 2025 that when underlying models fabricate contract addresses, partnerships, or price levels, those fabrications can propagate directly on-chain. Meta's chief AI scientist was cited in the same reporting as arguing the problem may require structural architectural redesigns rather than incremental training improvements.","heading":"Model-Specific Hallucination Rates and AI Agent Token Risks","severity":"critical","sources":[{"credibility":2,"name":"What DeepSeek-R1 Hallucinations Mean for 4 Crypto AI Agent Tokens — BeInCrypto","type":"news_article","url":"https://beincrypto.com/deepseek-r1-hallucination-crypto-ai-tokens/"}]},{"content":"AI-assisted smart contract auditing has grown significantly as a cost-reduction tool, with multiple security firms deploying LLM-powered audit pipelines. However, security researchers report that AI auditing tools generate hallucinated vulnerability reports — both false negatives, where dangerous code patterns are assessed as safe, and false positives, where safe code is flagged as vulnerable. The Zealynx blockchain security glossary notes that false positives waste resources on non-issues while false negatives allow real vulnerabilities to reach production. This is compounded by a documented finding that 100% of major crypto protocol hacks in 2024 hit audited projects, demonstrating that audit reports — whether produced by AI or humans — do not guarantee safety. Coinbase has developed an autonomous multi-phase AI auditing system called Frosty that uses sequential phased architecture to reduce hallucination rates, but acknowledges that reducing false positives remains a primary open challenge. Hacken security researchers separately noted that AI auditors frequently hallucinate code lines that do not exist, wasting reviewer time and creating false confidence in audit completeness.","heading":"Smart Contract Auditing: False Negatives and False Positives","severity":"high","sources":[{"credibility":2,"name":"AI Hallucination — Blockchain Security Glossary, Zealynx","type":"research","url":"https://www.zealynx.io/glossary/ai-hallucination"},{"credibility":2,"name":"Consumer Protection Tuesday: AI-Powered Smart Contract Auditing at Coinbase","type":"official","url":"https://www.coinbase.com/blog/consumer-protection-tuesday-ai-powered-smart-contract-auditing-at-coinbase"},{"credibility":2,"name":"Breaking Solidity at Scale: AI Smart Contract Auditing and the Workflow That Catches What AI Misses — Hacken","type":"research","url":"https://hackenproof.com/blog/for-hackers/ai-smart-contract-auditing-zakaria-hackenproof"}]},{"content":"Decentralized finance protocols depend on price oracles for liquidations, swaps, and collateral valuations. Research published on arXiv in July 2025 (arXiv 2507.02125) examining whether AI can solve the blockchain oracle problem found that LLM hallucinations within oracle contexts pose a specific risk: an AI price oracle might hallucinate market conditions, trading volumes, or liquidity depths, producing incorrect price feeds that automatically trigger on-chain actions. The immutable nature of blockchain records means that once a hallucinated price feed triggers a liquidation or swap, the transaction cannot be reversed. Zealynx's security glossary additionally flags prompt injection as a compounding attack vector: adversaries can deliberately craft inputs that induce hallucinations in AI oracle systems, potentially triggering erroneous liquidations or enabling market manipulation. Context poisoning — feeding models misleading information through compromised documentation or malicious forum posts that appear in Retrieval-Augmented Generation (RAG) pipelines — increases the probability that a model hallucinates conclusions beneficial to an attacker while those conclusions appear grounded in the poisoned context.","heading":"Oracle and DeFi Price Feed Risks","severity":"critical","sources":[{"credibility":2,"name":"Can Artificial Intelligence Solve the Blockchain Oracle Problem? (arXiv 2507.02125)","type":"research","url":"https://arxiv.org/pdf/2507.02125"},{"credibility":2,"name":"AI Hallucination — Blockchain Security Glossary, Zealynx","type":"research","url":"https://www.zealynx.io/glossary/ai-hallucination"}]},{"content":"In August 2025, CoinDesk and SentinelLABS documented a campaign in which over $1 million was stolen from crypto users through malicious smart contracts disguised as AI MEV trading bots. The campaign used AI-generated YouTube videos, aged accounts, and obfuscated Solidity code to instruct victims to deploy contracts and call a Start() function. The most prolific identified attacker address — 0x8725...6831 — accumulated 244.9 ETH (approximately $902,000), linked to a video with 387,000 views. The +30% profit shown in tutorial videos was fabricated through false balance return values or edited footage. While this incident involved deliberately malicious actors using AI content generation tools, it illustrates how AI-generated output creates a trust surface that users cannot reliably distinguish from legitimate information. Separately, the FBI's 2024 Internet Crime Report documented over $3.9 billion in crypto-related fraud losses in the U.S. alone, with the report including a first-ever dedicated AI section covering more than 22,000 complaints and approximately $893 million in AI-assisted fraud losses. Globally, crypto scam losses topped $10.7 billion in 2024.","heading":"Documented Financial Loss Incidents","severity":"critical","sources":[{"credibility":1,"name":"Weaponized Trading Bots Drain $1M From Crypto Users via AI-Generated YouTube Scam — CoinDesk","type":"news_article","url":"https://www.coindesk.com/tech/2025/08/07/weaponized-trading-bots-drain-usd1m-from-crypto-users-via-ai-generated-youtube-scam"},{"credibility":1,"name":"FBI: Crypto, AI Scams Drove Billions in Losses in 2025 — GovTech","type":"regulatory","url":"https://www.govtech.com/security/fbi-crypto-ai-scams-drove-billions-in-losses-in-2025"},{"credibility":2,"name":"Traders lose $1 million to malicious trading bot software — Web3 Is Going Great","type":"news_article","url":"https://www.web3isgoinggreat.com/?id=mev-bot-scam"}]},{"content":"A class of hallucination-enabled attack known as slopsquatting poses a distinct threat to blockchain developers. When LLMs generate code recommendations, they frequently suggest non-existent package names. Researchers coined the term slopsquatting (attributed to Seth Larson) to describe the practice of malicious actors registering these hallucinated names with malicious code. A 2025 study with over 200,000 code generation examples found that open-source models hallucinated fake package names at four times the rate of commercial models. USENIX Security 2025 research across 16 models and 576,000 code samples found that 38% of hallucinated packages were conflations of two real packages, 13% were typo variants, and 51% were pure fabrications. A notable documented example: security researcher Bar Lanyado of Lasso Security observed AI models consistently hallucinating a Python package called huggingface-cli — the non-existent short form of a real tool. The package accumulated over 30,000 authentic downloads in three months after a malicious actor registered it. For blockchain developers, this attack surface includes crypto-specific npm packages. A threat actor documented on dark web forums automated the creation of thousands of typosquatted packages targeting crypto libraries, using ChatGPT to generate plausible-sounding variants of real package names.","heading":"Slopsquatting: Hallucinated Package Names as a Supply Chain Attack Vector","severity":"high","sources":[{"credibility":2,"name":"AI Hallucinations and Slopsquatting: A Caution for Blockchain Devs — CCN","type":"news_article","url":"https://www.ccn.com/education/crypto/ai-hallucinations-slopsquatting-risk-for-blockchain-devs/"},{"credibility":2,"name":"Slopsquatting: When AI Agents Hallucinate Malicious Packages — Trend Micro","type":"research","url":"https://www.trendmicro.com/vinfo/us/security/news/cybercrime-and-digital-threats/slopsquatting-when-ai-agents-hallucinate-malicious-packages"},{"credibility":2,"name":"Package Hallucination: Impacts and Mitigation — Snyk","type":"research","url":"https://snyk.io/articles/package-hallucinations/"}]},{"content":"An April 2026 Fortune investigation highlighted what researchers from Microsoft Research, Columbia University, and Google DeepMind have termed the guarantee gap: the fundamental disconnect between the probabilistic reliability that AI safety techniques provide and the enforceable guarantees users require before delegating high-stakes financial tasks. The article presented scenarios in which an AI financial agent executing transactions misreads parameters, makes unauthorized leveraged bets, and produces realized losses — with no established mechanism for reimbursement or accountability. The research coalition proposed the Agentic Risk Standard (ARS), a layered settlement framework including escrow vaults releasing service fees only upon verified task delivery, collateral requirements for AI service providers, and third-party underwriting that prices the risk of AI failure and commits to user reimbursement. FINRA's 2026 Annual Regulatory Oversight Report, published in December 2025, included the first-ever dedicated section on generative AI, warning broker-dealers to develop specific procedures addressing hallucinations and bias, and to scrutinize AI agents that may act beyond the user's actual or intended scope. The SEC was noted as monitoring these developments.","heading":"Agentic AI and the Liability Gap","severity":"high","sources":[{"credibility":1,"name":"What do you do when your AI agent hallucinates with your money? — Fortune","type":"news_article","url":"https://fortune.com/2026/04/08/agent-hallucinations-protocol-money-financial-system-economy/"},{"credibility":1,"name":"GenAI: Continuing and Emerging Trends — FINRA 2026 Annual Regulatory Oversight Report","type":"regulatory","url":"https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai"},{"credibility":2,"name":"FINRA Cautions Broker/Dealers on Gen AI Hallucinations — Wealth Management","type":"news_article","url":"https://www.wealthmanagement.com/regulation-compliance/finra-cautions-broker-dealers-to-catch-hallucinations-when-using-gen-ai"}]},{"content":"Given the documented failure rates of AI agents in adversarial crypto environments, security researchers and regulators recommend a set of verification practices before acting on AI-generated crypto information. Users should independently verify any contract address provided by an AI tool against the official project website, block explorers (Etherscan, Solscan), and multiple independent sources before sending funds. Token addresses are particularly high-risk because a single hallucinated character in a 42-character address routes funds to an unrelated or malicious wallet. For smart contract security, audit reports produced by AI-only pipelines should not be treated as equivalent to human-reviewed audits; users should confirm that audit firms have manual review processes and published methodologies. Package names in developer tools recommended by AI coding assistants should be verified against official registries (npm, PyPI) before installation. For market data and protocol research, claims about TVL, APY, token supply, partnerships, or audit status should be cross-referenced with on-chain data from block explorers and official protocol documentation rather than relying solely on AI-generated summaries. Multi-source platforms that aggregate and cross-reference claims reduce but do not eliminate hallucination risk, as evidenced by research into RAG-based mitigation strategies.","heading":"Verification Best Practices for Crypto Users","severity":"medium","sources":[{"credibility":2,"name":"AI Hallucination — Blockchain Security Glossary, Zealynx","type":"research","url":"https://www.zealynx.io/glossary/ai-hallucination"},{"credibility":2,"name":"The two faces of AI in crypto: Threats, opportunities and what Elliptic is doing about it — Elliptic","type":"research","url":"https://www.elliptic.co/blog/the-two-faces-of-ai"},{"credibility":2,"name":"When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets (arXiv 2510.00332)","type":"research","url":"https://arxiv.org/abs/2510.00332"}]},{"content":"The hallucination risk documented across academic literature and regulatory guidance directly informs the design requirements of trust intelligence platforms operating in the crypto space. Elliptic, a blockchain analytics firm, publicly disclosed that in building its AI-powered compliance copilot, its teams required strong assurances against hallucination before deployment, implementing MLflow tracing to make every model response traceable and auditable. The firm described the challenge as ensuring that outputs would not hallucinate, omit key context, or rely on opaque reasoning when feeding into formal investigations and regulatory reports. AVOID.NET addresses this risk category through a multi-agent verification pipeline in which a research agent's initial findings are independently reviewed by a separate fact-checking agent and adjudicated by a judge agent, with all accepted evidence anchored to on-chain records via Solana SPL Memo. This approach reflects the academic consensus — documented in arXiv 2510.00332 and 2512.03107 — that multi-step human-in-the-loop validation and cross-source grounding are the most reliable available mitigations for LLM hallucinations in high-stakes financial contexts. No single-agent AI output should be treated as a definitive trust determination without independent verification.","heading":"Trust Intelligence Implications and Multi-Source Verification","severity":"medium","sources":[{"credibility":2,"name":"The two faces of AI in crypto: Threats, opportunities and what Elliptic is doing about it — Elliptic","type":"research","url":"https://www.elliptic.co/blog/the-two-faces-of-ai"},{"credibility":2,"name":"Protect Crypto Assets with Governed AI — Databricks (Elliptic case study)","type":"other","url":"https://www.databricks.com/customers/elliptic"},{"credibility":2,"name":"Tool Receipts, Not Zero-Knowledge Proofs: Practical Hallucination Detection for AI Agents (arXiv 2603.10060)","type":"research","url":"https://arxiv.org/pdf/2603.10060"}]}],"sources_used":[{"credibility":2,"name":"When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets (arXiv 2510.00332)","type":"research","url":"https://arxiv.org/abs/2510.00332"},{"credibility":2,"name":"Detecting AI Hallucinations in Finance: An Information-Theoretic Method Cuts Hallucination Rate by 92% (arXiv 2512.03107)","type":"research","url":"https://arxiv.org/abs/2512.03107"},{"credibility":2,"name":"Agentic AI and Hallucinations (arXiv 2507.19183)","type":"research","url":"https://arxiv.org/pdf/2507.19183"},{"credibility":2,"name":"Can Artificial Intelligence Solve the Blockchain Oracle Problem? (arXiv 2507.02125)","type":"research","url":"https://arxiv.org/pdf/2507.02125"},{"credibility":2,"name":"Tool Receipts, Not Zero-Knowledge Proofs: Practical Hallucination Detection for AI Agents (arXiv 2603.10060)","type":"research","url":"https://arxiv.org/pdf/2603.10060"},{"credibility":2,"name":"Giving AI Agents Access to Cryptocurrency and Smart Contracts Creates New Vectors of AI Harm (arXiv 2507.08249)","type":"research","url":"https://arxiv.org/html/2507.08249v1"},{"credibility":2,"name":"AI Hallucination — Blockchain Security Glossary, Zealynx","type":"research","url":"https://www.zealynx.io/glossary/ai-hallucination"},{"credibility":2,"name":"What DeepSeek-R1 Hallucinations Mean for 4 Crypto AI Agent Tokens — BeInCrypto","type":"news_article","url":"https://beincrypto.com/deepseek-r1-hallucination-crypto-ai-tokens/"},{"credibility":1,"name":"Weaponized Trading Bots Drain $1M From Crypto Users via AI-Generated YouTube Scam — CoinDesk","type":"news_article","url":"https://www.coindesk.com/tech/2025/08/07/weaponized-trading-bots-drain-usd1m-from-crypto-users-via-ai-generated-youtube-scam"},{"credibility":1,"name":"FBI: Crypto, AI Scams Drove Billions in Losses in 2025 — GovTech","type":"regulatory","url":"https://www.govtech.com/security/fbi-crypto-ai-scams-drove-billions-in-losses-in-2025"},{"credibility":2,"name":"Traders lose $1 million to malicious trading bot software — Web3 Is Going Great","type":"news_article","url":"https://www.web3isgoinggreat.com/?id=mev-bot-scam"},{"credibility":1,"name":"What do you do when your AI agent hallucinates with your money? — Fortune","type":"news_article","url":"https://fortune.com/2026/04/08/agent-hallucinations-protocol-money-financial-system-economy/"},{"credibility":1,"name":"GenAI: Continuing and Emerging Trends — FINRA 2026 Annual Regulatory Oversight Report","type":"regulatory","url":"https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai"},{"credibility":2,"name":"FINRA Cautions Broker/Dealers on Gen AI Hallucinations — Wealth Management","type":"news_article","url":"https://www.wealthmanagement.com/regulation-compliance/finra-cautions-broker-dealers-to-catch-hallucinations-when-using-gen-ai"},{"credibility":2,"name":"AI made crypto scams far more dangerous — Help Net Security","type":"news_article","url":"https://www.helpnetsecurity.com/2025/09/18/ai-crypto-scams-dangerous/"},{"credibility":2,"name":"AI Hallucinations and Slopsquatting: A Caution for Blockchain Devs — CCN","type":"news_article","url":"https://www.ccn.com/education/crypto/ai-hallucinations-slopsquatting-risk-for-blockchain-devs/"},{"credibility":2,"name":"Slopsquatting: When AI Agents Hallucinate Malicious Packages — Trend Micro","type":"research","url":"https://www.trendmicro.com/vinfo/us/security/news/cybercrime-and-digital-threats/slopsquatting-when-ai-agents-hallucinate-malicious-packages"},{"credibility":2,"name":"Package Hallucination: Impacts and Mitigation — Snyk","type":"research","url":"https://snyk.io/articles/package-hallucinations/"},{"credibility":2,"name":"The two faces of AI in crypto: Threats, opportunities and what Elliptic is doing about it — Elliptic","type":"research","url":"https://www.elliptic.co/blog/the-two-faces-of-ai"},{"credibility":2,"name":"Protect Crypto Assets with Governed AI — Databricks (Elliptic case study)","type":"other","url":"https://www.databricks.com/customers/elliptic"},{"credibility":2,"name":"Consumer Protection Tuesday: AI-Powered Smart Contract Auditing at Coinbase","type":"official","url":"https://www.coinbase.com/blog/consumer-protection-tuesday-ai-powered-smart-contract-auditing-at-coinbase"},{"credibility":2,"name":"Breaking Solidity at Scale: AI Smart Contract Auditing — Hacken","type":"research","url":"https://hackenproof.com/blog/for-hackers/ai-smart-contract-auditing-zakaria-hackenproof"}],"summary":"AI hallucinations — outputs generated by large language models that appear factually credible but are fabricated — present a distinct and growing threat category in cryptocurrency markets. When AI-powered tools such as trading bots, smart contract auditors, research platforms, and oracle providers produce hallucinated data, the consequences can propagate directly on-chain as irreversible financial transactions. Academic benchmarks, regulatory warnings, and documented incidents collectively confirm that current AI agent accuracy in adversarial crypto environments falls well short of the reliability threshold required for autonomous financial decisions.","timeline":[{"date":"2024-01-01","event":"$30 billion lost to exploits across cryptocurrency markets during the 2024 calendar year, the context cited by academic researchers benchmarking AI agent reliability in adversarial crypto environments.","source":"When Hallucination Costs Millions (arXiv 2510.00332)","source_url":"https://arxiv.org/abs/2510.00332"},{"date":"2024-03-01","event":"Security researcher Bar Lanyado of Lasso Security observed AI models consistently hallucinating a non-existent Python package 'huggingface-cli'; after a malicious actor registered the name, it accumulated over 30,000 downloads in three months — an early documented instance of slopsquatting.","source":"Slopsquatting: When AI Agents Hallucinate Malicious Packages — Trend Micro","source_url":"https://www.trendmicro.com/vinfo/us/security/news/cybercrime-and-digital-threats/slopsquatting-when-ai-agents-hallucinate-malicious-packages"},{"date":"2025-01-20","event":"DeepSeek-R1 released. Independent benchmarking by Vectara using HHEM 2.1 found DeepSeek-R1 exhibited a 14.3% hallucination rate, approximately four times higher than DeepSeek-V3 (3.9%), raising concerns for the hundreds of AI agent tokens relying on reasoning-style LLMs for on-chain execution.","source":"What DeepSeek-R1 Hallucinations Mean for 4 Crypto AI Agent Tokens — BeInCrypto","source_url":"https://beincrypto.com/deepseek-r1-hallucination-crypto-ai-tokens/"},{"date":"2025-08-07","event":"CoinDesk and SentinelLABS published findings documenting over $1 million stolen through AI-generated YouTube videos promoting malicious smart contracts disguised as MEV trading bots. The most successful attacker address accumulated 244.9 ETH (approximately $902,000).","source":"Weaponized Trading Bots Drain $1M From Crypto Users via AI-Generated YouTube Scam — CoinDesk","source_url":"https://www.coindesk.com/tech/2025/08/07/weaponized-trading-bots-drain-usd1m-from-crypto-users-via-ai-generated-youtube-scam"},{"date":"2025-09-18","event":"Help Net Security reported that generative AI-assisted crypto scams rose 456% between mid-2024 and mid-2025, with AI enabling more convincing social engineering, deepfake impersonation campaigns, and automated phishing at scale.","source":"AI made crypto scams far more dangerous — Help Net Security","source_url":"https://www.helpnetsecurity.com/2025/09/18/ai-crypto-scams-dangerous/"},{"date":"2025-10-01","event":"Researchers published 'When Hallucination Costs Millions: Benchmarking AI Agents in High-Stakes Adversarial Financial Markets' (arXiv 2510.00332), documenting that 17 evaluated AI models achieved only 28% accuracy without tools and 67.4% with tools on crypto tasks, far below the 80% human baseline.","source":"When Hallucination Costs Millions (arXiv 2510.00332)","source_url":"https://arxiv.org/abs/2510.00332"},{"date":"2025-12-09","event":"FINRA published its 2026 Annual Regulatory Oversight Report, marking the first time FINRA included a dedicated section on generative AI, warning broker-dealers to develop specific procedures targeting hallucinations and to scrutinize AI agents acting beyond user intent.","source":"GenAI: Continuing and Emerging Trends — FINRA 2026 Annual Regulatory Oversight Report","source_url":"https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai"},{"date":"2025-12-15","event":"Researchers published 'Detecting AI Hallucinations in Finance: An Information-Theoretic Method Cuts Hallucination Rate by 92%' (arXiv 2512.03107), proposing the ECLIPSE framework as a mitigation approach for financial LLM deployments.","source":"Detecting AI Hallucinations in Finance (arXiv 2512.03107)","source_url":"https://arxiv.org/abs/2512.03107"},{"date":"2026-04-08","event":"Fortune published an investigation into the guarantee gap facing AI financial agents, documenting a proposed Agentic Risk Standard (ARS) from researchers at Microsoft Research, Columbia University, and Google DeepMind as a response to the absence of liability frameworks for hallucination-induced financial losses.","source":"What do you do when your AI agent hallucinates with your money? — Fortune","source_url":"https://fortune.com/2026/04/08/agent-hallucinations-protocol-money-financial-system-economy/"}]},"v":1}

   Verify offline (run on your own machine)

   python -m src.verify_decision 6447cd48-15ed-4cd2-9278-97b401fe198acopy