By: E.J. Yerzak, Managing Director
It was only a matter of time. Artificial intelligence use by financial institutions has exploded in 2026. The more advanced reasoning and capability within these enterprise AI tools comes with a cost – either in the form of API charges per million tokens (MTok) used or baked into platform subscription charges that may include usage throttling. AI tools bring powerful computing power to front and back-office operations and portfolio management processes. And access to that computing power – as well as the sensitive data it is increasingly using – has now become a very attractive target for hackers.
We have seen this play out before, albeit in a different context. When cryptocurrency started gaining traction, individuals found they could use computing power to mine new Bitcoin. Significant computer processing power is required to solve the cryptography to create brand new Bitcoin. Hackers sought to short-circuit the process by using “cryptojacking” to gain unauthorized access to organizations’ computers, enabling them to direct that computing power towards cryptocurrency mining, often behind the scenes without users knowing why their computers are suddenly more sluggish.
Now the battlefield has shifted. In what is now being labeled “LLMjacking,” hackers have pivoted towards theft of corporate API keys and credentials to the growing number of AI platforms such companies have rolled out over the past year. It’s a classic two birds / one stone approach with the potential for a massive payoff for hackers. By compromising the API keys, malicious actors can not only run voluminous AI queries for their own purposes under a license your firm is paying for, but also to mine the confidential data increasingly being used in or accessible to enterprise AI tools.
On August 31, 2026, a nonprofit called METR disclosed that it experienced two security incidents this year. According to METR, one of those incidents involved hackers gaining access to the API key for its AI models, enabling the hackers to use up AI search credits worth approximately $600,000. The spike in usage went undetected for various reasons, including usage monitoring gaps. While no sensitive data other than perhaps the API key appears to have been compromised , the incident highlights the growing importance of strong AI governance practices.
Financial firms should take note. Periodic risk assessments are important not only to keep compliance and cybersecurity programs up to speed, but also to protect the business against an ever-evolving risk environment. Compromised data doesn’t have to be the type that requires breach notice to cause considerable damage to a firm. Internal documents alone could bring down a firm if their exposure reveals information about potential investments or planned strategies.
What can firms do to protect themselves? Establishing strong AI governance is a great place to start. Consider implementing the following controls as feasible.
- Implement controls around secure storage for credentials and API keys for AI environments.
- The principle of least privilege still applies. If an AI tool does not need access to all of your firm’s data, consider restricting its access. In the METR incident, hackers were able to trick an AI agent into disclosing an API key that it had access to.
- Establish usage benchmarks. The more granular the benchmarks, the better. Organizations who have a sense of their normal use are best positioned to detect anomalies in such use.
- Continuously monitor for unexpected network and application activity. Unexpected privilege escalation should ideally trip an alarm.
- Consider your attack surface holistically. Every endpoint poses a way into your organization.
Financial institutions who undertake periodic risk assessments are better able to stay abreast of these emerging threats. And the good news is that firms don’t need to rush out and adopt a standalone LLMjacking policy to solve this. The measures outlined above – keeping credentials secure, the principle of least privilege, monitoring for anomalous activity – are all components of a reasonably designed program already. What needs to continuously evolve is the risk assessment process and the implementation of controls to tackle these new risks.
Remember, compliance and cybersecurity programs were never intended to be approached with a “set it and forget it” mentality. Even the SEC’s Compliance Programs Rule requires reevaluating the effectiveness of controls at least annually to determine if they are still reasonably designed.
Adaptability is the key to effective risk management. As the risks evolve, so, too, must we evolve our programs. After all, who would have ever predicted that CCOs would have to learn about a term like “LLMjacking”?