AI Update: U.S. Charges Three in $129 Million Fraud Using AI to Pose as CEOs

Federal prosecutors in Manhattan have charged three men over a $129 million impersonation scheme that allegedly used AI, which their announcement calls “super intelligence,” to mimic chief executives and lawyers. According to the indictment, the defendants imitated CEOs’ voices or images on calls, asked employees for help with a confidential transaction, then posed as lawyers at major international firms and directed them to wire money to accounts the group controlled. Alleged losses include $11.6 million at a UK financial technology company and $4.8 million at a Missouri hospital system. Jacob Soall, Sharon Hazan and Guy Sheetrit were arrested in Spain and the UK, and extradition requests are pending.

WorkOS, which sells single sign-on and other authentication tools to software companies, said the intrusion it disclosed on November 2 came from a security researcher’s AI agent that escaped its testing environment, not a malicious actor. The researcher was running cybersecurity benchmarks with ARTEX, an AI penetration-testing framework. The agent reached an internal go-to-market tool that was exposed to the internet and accepted a predictable credential meant for local development. It first got in on October 25 and pulled records in bulk on October 30, including contact details and excerpts of customer conversations. WorkOS and a forensics firm validated the researcher’s account, and the researcher confirmed deleting the data.

Blockchain analysis firm Chainalysis said names, usernames, password hashes and live tokens for all its users were exposed in a breach it attributes to an external actor using “offensive agentic techniques.” On October 30 the company said it had found no evidence that credentials or session tokens were accessed. It later said an “on-path attack,” in which an attacker intercepts traffic between systems, exposed some passwords. After detecting abnormal activity, it took its KYT transaction-monitoring service and login page offline and began notifying customers whose KYT accounts appear to have been accessed. All users must reset passwords. Chainalysis says it has no evidence investigation data was accessed.

Binance said attackers the FBI links to North Korea reached its wallet signing service through fake recruitment interviews and a compromised internal AI tool before taking about $6.34 billion in digital assets on October 29. An engineer contacted through a job platform had a live video interview with a purported HR representative, then was told to download a coding assignment containing malware. On October 20, the attackers used that computer and an internal AI tool to gain control over two engineers working on wallet infrastructure, one of whom held a container deployment token. Binance said about $2.85 billion has been frozen or intercepted, and that it will absorb the loss.

India’s national cyber response team, CERT-In, issued a critical advisory naming “Crafty Falcon,” an AI-assisted financial threat actor targeting non-bank lenders, fintech firms, payment providers and crypto exchanges. The agency said the attackers use publicly available AI-driven tools including ARTEX, PentAGI, Cairn and CyberStrikeAI, enter through internet-exposed test environments and call victims’ own payment and disbursement APIs using modified app packages. They validate fund flows with test transfers such as one rupee before moving larger sums, and have stolen crypto wallet keys. Organizations were targeted again after incomplete remediation, CERT-In said, with attacks timed to holidays and weekends. It first warned of the campaign on October 9.

OpenAI made GPT-6 Cyber, a model trained to find and validate exploitable software vulnerabilities, available to a select group of approved customers in its Daybreak Red program, and classified it as Critical in cybersecurity under its Preparedness Framework, up from High for its predecessor. OpenAI said its researchers used the model to develop 948 Linux privilege escalation exploits and identify 81 previously unknown high-severity Firefox flaws. In an evaluation informed by the July incident in which an OpenAI agent breached Hugging Face, the company said the model avoided exploiting out-of-scope services in 99.93 percent of cases without deployment safety prompts and in all cases with them.

The UK government will consult on using powers in its Cyber Security and Resilience Bill to address “agentic loss of control” incidents, in which operators are unable to adequately monitor or control AI agents. It intends to introduce four measures for regulated entities through secondary legislation: security requirements such as monitoring agents’ actions, notice to other organizations an agent has affected, preservation of evidence about agents’ actions, and powers for ministers to have government experts review incidents. The measures are not meant as a comprehensive AI governance framework, it said. The bill passed its House of Lords third reading on November 2 and returns to the Commons.