When machines attack: Frontier AI cyber threats and policy responses in the financial sector
Planned steps: Identify frontier AI cyber threats, quantify the 2025 attack increase, and summarise policy responses. This report details an 89% rise in AI-enabled cyber attacks, highlighting frontier models’ ability to autonomously exploit vulnerabilities. Financial authorities are responding by reinforcing existing frameworks, prioritising accelerated patching, and enhancing governance.
Please login or join for free to read more.
OVERVIEW
Section 1 – Introduction
Frontier artificial intelligence (AI) models are rapidly reshaping the cyber threat landscape. Unlike previous generations that focused on discrete tasks, frontier AI increasingly demonstrates the capacity to autonomously identify software vulnerabilities, develop attack methods at scale, and execute multi-step cyber operations. These developments present a significant challenge to the operational resilience of financial institutions, which rely on interconnected digital infrastructures and third-party technological providers. While frontier AI offers opportunities to strengthen cyber defence, it also drastically shortens the time available for organisations to respond to incidents.
Section 2 – AI and cyber capabilities
The cyber security landscape has reached a new phase where AI-enabled attacks increased by 89% in 2025. The speed of these intrusions is accelerating, with the average eCrime breakout time—the time it takes for an attacker to move laterally within a network—falling to 29 minutes in 2025. This represents a 65% increase in speed from the previous year, significantly narrowing the window for defenders to detect and respond to threats effectively.
The use of AI in cyber attacks
Malicious actors were misusing AI models even before the release of frontier versions. A study by Anthropic mapped 13,873 actions to 14 tactics in the MITRE ATT&CK framework. The research found that the most prevalent tactic enabled by AI was defence impairment at 16.17%, which involves compromising security controls to avoid detection. Other significant uses included resource development at 13.27% and collection at 9.45%. AI is increasingly utilised once attackers are inside a system, suggesting that models are becoming more capable of the complex reasoning required for lateral movement and data exfiltration.
Case study: AI-assisted breach of Mexico’s government infrastructure
Between December 2025 and February 2026, a single threat actor used Anthropic’s Claude Code and OpenAI’s GPT-4.1 to breach nine Mexican government agencies. The AI served as a primary operational tool, automating reconnaissance and privilege escalation to exfiltrate over 150 gigabytes of sensitive data belonging to up to 195 million individuals. The campaign compressed attack timelines to within standard detection windows, enabling a single operator to process reconnaissance data that would typically require a team of analysts. However, many exploited vulnerabilities could have been mitigated through standard security controls such as patching and credential rotation.
Frontier AI cyber capabilities
The release of models like Claude Mythos and GPT-5.5 marks a step change in offensive potential. These models excel at autonomous vulnerability discovery and exploitation. Claude Mythos reportedly identified over 10,000 high-severity vulnerabilities across systemically important software. In performance evaluations, GPT-5.5 achieved a 71.4% pass rate on expert-level cyber tasks. Furthermore, the UK AI Security Institute (AISI) estimates that the length of cyber tasks these models can complete reliably is doubling every 4.7 months. Testing on simulated cyber ranges showed that frontier models are now capable of autonomously completing multi-step intrusions, such as the 32-step “The Last Ones” corporate network intrusion.
Quantitative evidence highlights the efficiency of these models; Claude Mythos Preview produced 157 working exploits (17% success rate) in ExploitGym benchmarks, compared to just 15 (2%) for Claude Opus 4.6. Additionally, unpatched software has become the leading initial access vector for breaches, accounting for 31% of incidents, while credential abuse has dropped to 13%.
Section 3 – Financial authorities’ policy responses
Financial authorities are converging on a pragmatic response by reinforcing existing cyber risk management and operational resilience frameworks rather than introducing AI-specific regimes. Policy responses emphasise board-level AI literacy, the integration of AI-enabled risks into risk appetite frameworks, and a clear decision-making hierarchy. Authorities in jurisdictions like the US, UK, and EU are establishing collaborative task forces and clearing houses to coordinate vulnerability discovery and patching across critical infrastructure.
Key recommendations include accelerating patching cycles, as traditional scheduled cycles are insufficient against frontier AI speeds. For instance, APRA and ASIC have urged firms to strengthen cyber resilience fundamentals and reduce attack surfaces. Hong Kong’s HKMA and Japan’s FSA have promoted structured information-sharing and public-private working groups to exchange threat intelligence. There is a heightened focus on third-party dependencies, as reliance on a few common frontier AI providers introduces concentration and sovereign access risks.
Section 4 – Concluding remarks
Frontier AI represents a critical evolution in cyber risk due to its speed, scale, and autonomy. While the underlying attack techniques remain familiar, AI significantly reduces the expertise needed to execute them. Financial authorities must continue to prioritise basic cyber hygiene, regular reassessment of controls, and realistic testing. Closing gaps in supervisory expertise and fostering international coordination will be essential to keeping pace with a rapidly changing threat environment where vulnerability discovery may soon be measured in hours rather than weeks.