AI-assisted hacking is moving deeper into the exploitation process. The Federal Reserve now warns that frontier AI models can reportedly find unknown software vulnerabilities, devise exploits, and chain multiple flaws into attacks executed at machine speed.
The Sept. 4 report also acknowledges that the Federal Reserve Board’s information-security program received a level-3, “not effective” rating in the Office of Inspector General’s 2025 FISMA assessment. The Fed says subsequent technical assessments found the weaknesses were primarily nontechnical and concentrated in governance.
Shared cloud, software, and service-provider dependencies raise the stakes. The Fed’s 2026 Cybersecurity and Financial System Resilience Report says incidents at critical third parties can affect multiple organizations, while faster AI-assisted exploitation could narrow the time available to remediate flaws before attackers can use them.
AI exploit chains shrink the defense window
The Fed attributes the capability to leading researchers rather than its own testing. Those researchers report that frontier models can identify previously undiscovered vulnerabilities, create methods to exploit them, and connect flaws into attack chains that can run at machine speed.
The same technology is also changing defensive security. Microsoft has said AI is helping its researchers find more software vulnerabilities, illustrating how faster discovery can benefit defenders while increasing pressure to assess and patch flaws quickly.
The Fed says controlled access to powerful models could give defenders an opportunity to identify and fix vulnerabilities first. The same capability could provide malicious actors with vulnerability intelligence or offensive tools before weaknesses are fully addressed across financial institutions and their technology providers.
The threat remains a capability warning, not evidence of a bank breach. The Fed, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation said they had not observed material impacts on the financial sector from the threat environment described in the report.
Frontier-model development nevertheless shows how quickly that capability threshold is moving. OpenAI’s latest Astra model reached the company’s highest cybersecurity capability tier, which OpenAI says covers systems capable of finding unknown vulnerabilities and developing exploits against well-protected targets.
The Fed Board has a governance gap of its own
The Board’s October 2025 OIG audit provides the clearest picture of its governance problem. The information-security program fell from level 4, “managed and measurable,” to level 3, “consistently implemented,” below the level considered effective under the federal maturity framework.
The OIG said governance challenges stemming from the Board’s decentralized IT model contributed to the decline, alongside weaknesses involving cybersecurity profiles, mobile-device security, and classification of confidential supervisory information. Eleven of the Board’s 13 divisions had embedded IT groups. Only 48 of 88 full-time cybersecurity personnel reported to the CIO.
The Fed’s September report says the Board had identified governance weaknesses before the audit and is implementing a new information-security operating model. Later third-party assessments, according to the Fed, found the remaining weaknesses were primarily nontechnical and concentrated in governance, while other security domains improved or remained stable.
The finding mirrors a broader enterprise problem: organizations are deploying AI faster than many are formalizing controls around it. A July survey found that enterprise AI governance programs lagged behind deployment, with only about half of surveyed organizations reporting formal governance programs despite widespread AI use.
What eWeek found: The Fed’s AI threat model has shifted from assisted attacks to automated exploit chains
The Fed’s own reporting shows how quickly its AI threat model has changed. Earlier cybersecurity reports focused largely on AI helping attackers automate reconnaissance, phishing, and social engineering. By 2026, the regulator is highlighting models that can find previously unknown vulnerabilities, build exploits, and combine them into machine-speed attack chains.
That shift moves the enterprise problem from AI-assisted attacks toward increasingly automated exploitation. Organizations still depend on human-led patching, vendor coordination, access governance, and incident response, making fragmented security ownership more consequential as the technical attack cycle accelerates.
Read more: As frontier models move deeper into security operations, OpenAI Astra and CrowdStrike SafeMind show how general-purpose and purpose-built cyber AI are beginning to converge across the enterprise security stack.
Want to learn more AI tips, tricks, and prompting techniques? eWeek readers get free 7-day access to The Neuron Academy, our practical learning platform designed to help professionals use AI more confidently at work. Browse all lessons →