AI is starting to automate parts of cyberattacks, but reports of machine-led hacking do not prove attackers can run entire operations without people. A warning attributed to Felix Fan, together with findings from Google and India’s CERT-In, points to a more immediate problem: security teams may be too slow to spot and contain attacks outside business hours.
- Google says it observed AI frameworks automating vulnerability discovery and exploitation.
- The findings do not show how common fully autonomous attacks are.
- Self-custody gives users control, but no guarantee against theft or recovery.
- Trust Wallet’s extension incident shows the difference.
AI can automate tasks without running the whole attack
In a post on X dated Nov. 18, 2025, Felix Fan warned that banks and other large organizations may be unprepared for attacks that move faster than conventional security teams. He pointed to the risk of companies holding substantial customer funds while relying on IT departments that mainly operate on weekdays. Fan also said he would personally rather keep funds in Trust Wallet than in a bank.
Those remarks are a warning and a personal preference. They do not show that every bank is poorly secured or that a crypto wallet is safer. The concrete concern is whether organizations can detect and respond to attacks when monitoring or staffing is reduced.
Google’s Threat Intelligence Group (GTIG), in its Nov. 16, 2025 AI Threat Tracker, said it had observed threat actors using large language model (LLM) frameworks to automate vulnerability discovery and exploitation without human intervention “for the first time.” LLMs are AI systems that process and generate language. Frameworks can connect them to tools and tasks, allowing them to carry out sequences of actions.
Google also tracked a threat actor using an automated framework to conduct reconnaissance and scan organizations across Asia between late August and early October. Reconnaissance is the early work of gathering information about potential targets. A vulnerability is a flaw that can expose software or systems to unauthorized access. A patch is an update designed to fix one. Organizations can remain exposed if they have not installed an available patch.
“Without human intervention” is significant, but it does not necessarily mean AI chose the target, set the objective and carried out every stage of an operation. Automation can handle one task while people remain involved elsewhere. Google’s observation is a warning about capability, not a measure of how widespread successful autonomous attacks are.
Security coverage cannot stop at closing time
India’s Computer Emergency Response Team, CERT-In, warned on Nov. 4, 2025, of escalating AI-assisted attacks targeting financial technology companies, digital lenders, payment service providers and cryptocurrency exchanges. CERT-In identified a threat actor it designated CRAFTY FALCON and said some organizations had suffered repeat intrusions despite earlier security recommendations.
The agency said attackers deliberately scheduled operations during weekends, holidays and other periods when monitoring was reduced. It recommended continuous monitoring, extra security staffing during those periods, faster software updates and stronger transaction controls. These are basic measures, but automation makes neglecting them riskier. A vulnerability will not wait for the office to reopen.
Google’s chief analyst, John Hultquist, has argued that defenders will need “agentic defense, ” or automated and AI-enabled systems that can help monitor threats and respond. That is a proposed direction, not proof that AI defenses already work reliably against autonomous attacks. Claims about AI security products deserve the same scrutiny as claims about AI attackers.
Self-custody changes the risk; it does not remove it
Fan’s preference for Trust Wallet raises a separate question: does holding crypto yourself make funds safer than keeping them at a bank? Self-custody means users control their assets through private keys or a recovery phrase instead of relying on a custodian. It can reduce dependence on a financial intermediary, but users become responsible for protecting those credentials and devices. If keys are stolen or exposed, there may be no central authority able to reverse the transaction.
A Trust Wallet browser-extension incident offers a concrete reminder. Trust Wallet says a compromised version 2.68 of its Chrome extension affected 2, 520 wallet addresses. The company reported approximately $8.5 million in assets associated with 17 attacker-controlled addresses, while cautioning that those addresses also drained wallets it did not associate with Trust Wallet or the incident. That means the $8.5 million figure is not a confirmed total loss by Trust Wallet users.
Trust Wallet says it chose to reimburse affected users voluntarily and is reviewing claims case by case. This is a company-led process, not automatic compensation or a general guarantee against future losses. The company’s preliminary assessment linked the incident to the wider Sha1-Hulud activity, but said its investigation into the precise attack path was still ongoing.
Eligible deposits at insured U.S. banks may receive FDIC protection under applicable rules and limits. That protection does not cover every kind of bank-related loss, and self-custodied cryptocurrency is not an insured bank deposit. Banks and crypto come with different protections and responsibilities.
Key questions and answers
-
What did Google say it observed?
Google’s Threat Intelligence Group said it had observed attackers using LLM frameworks to automate vulnerability discovery and exploitation without human intervention. The finding does not show how prevalent such activity is or prove that entire attacks were run autonomously.
-
What do the Trust Wallet incident figures mean?
Trust Wallet reported 2, 520 affected addresses and about $8.5 million in assets associated with 17 attacker-controlled addresses. The company said some of those addresses also drained wallets it did not link to the incident, so the figure is not a confirmed total of Trust Wallet users’ losses.
-
Does self-custody come with deposit insurance?
No. Eligible deposits at insured U.S. banks may qualify for FDIC protection under applicable rules and limits. Self-custodied crypto is not an insured bank deposit. Users who hold their own assets are also directly responsible for protecting their keys and recovery phrases.
Automation makes slow patching and limited monitoring more dangerous, but it does not make every attacker unstoppable or every AI defense effective. Self-custody offers control, not immunity. Removing a custodian also removes a potential safety net, and recovery is never guaranteed. India’s warning on AI-driven cyberattacks shows why security teams need to be ready outside business hours.