Denis Mandich, CTO of Qrypt, a quantum cybersecurity company, founding member of the Quantum Economic Development Consortium and CQT.
With the flood of generative AI (GenAI) content, many people now believe the often-recited misconception that “no one understands quantum mechanics” now applies to GenAI, believing that no one knows how GenAI works. This conflation of a pithy quote by physicist Richard Feynman is untrue, even though it has been relayed as fact by many industry leaders.
This perspective is forgivable because AI is a new and empirically advancing field, unlike quantum physics, a science established well over a century ago. LLM hallucinations are a pick-your-euphemism for wrong, incompetent, lying, incoherent or dumb.
It’s a good reminder of how far we are from artificial general intelligence (AGI), regardless of definition and how LLMs’ statistical sequencing of words is uncorrelated with intelligence.
Even amid this confusion, cybersecurity marketing is plagued by attaching the latest buzzword to every product, with little substantive innovation or differentiation, rendering it meaningless. Anomaly detection, zero trust and XYZ detection response have all been rebranded as AI.
Regardless of what constitutes AI, some version of near-fully automated systems is essential today. Humans are inefficient and incapable of keeping up with today’s cyberattacks.
This is why the issue of AI cybersecurity is difficult to address, much less fix. Would the best AI-powered anomaly detection tools have flagged the errors generated by AI tools? No, because they are based on the vast training data scraped from the internet, much of it wrong. AI is incapable of intelligent decision-making in novel cyberattacks because, by definition, it has not learned anything that was not already in its training data.
Cybersecurity’s AI Challenge
Even with countless machine learning and AI models, datasets and apps in the open-source communities, AI can only grow incrementally without a bigger quantum computer. The next leaps forward will require a hybrid quantum solution. Until then, there are difficult-to-quantify questions about AI.
The ability to advance AI fast enough will create cybersecurity problems because new attacks are truly insidious. The goals are transitioning from stealing data and doing reconnaissance to modifying the AI control systems and training data. One example is the Hugging Face security breach, where access keys and tokens to authenticated services were compromised. No information about the attackers and methods has been disclosed.
There are similarities here to the Storm-0558 compromise of Microsoft’s master signing key, which allowed hackers to “gain full access to essentially any Exchange Online account anywhere in the world,” according to the Cyber Safety Review Board. Setting aside the fact that a god-like key should not exist, a year has passed, and we still don’t know how it happened.
How AI Can Bridge The Gap To Quantum Computing
It is false to claim that AI LLMs are not well understood. The premise is absurd: It would mean that AI-powered cybersecurity companies want you to use their tools even though they don’t know how or why they work.
That said, safety and security measures are not built into AI systems, nor is it rational to expect an AI expert to have the skills to understand how to do so. Cybersecurity company products fail regularly, so it would be unrealistic to expect an AI developer to anticipate the sophisticated intrusion methods used by nation-states.
That doesn’t mean the industry can’t work to bridge the gap. There was outrage recently about OpenAI appointing former NSA director Paul Nakasone to its board. Despite concerns, this should be considered a rational step in the right direction. Few people can claim Nakasone’s knowledge of the adversary’s capabilities and nefarious AI objectives.
As we further automate and add AI decision-making to our critical infrastructure, we need extraordinary security guarantees, not amazing AI functionality with dubious safety measures. These must include third-party evaluation and validation of AI tools used for cybersecurity, which must have some government regulation attached.
Virtually every important industry has compliance requirements with consequences ranging from financial to industry sanctions and decertification. AI cannot be an exception when entrusted with so much responsibility. The demand must be driven by corporate adopters of any technology, not the technology producers and advocates.
Reaching Quantum Computing
While the fears over deep fakes grab headlines, quiet versions of this methodology by external adversaries are the real threat.
Think Stuxnet, where fake sensor data tricked an automated nuclear technology control system into self-destruction. The zero-day Log4j also scored 10 out of 10 in CVSS severity because it allowed remote control and code execution of vulnerable systems.
There is no equivalent rating system when compromised AI already has control by design. Triggering a desired response may only require certain conditions to be met from data or sensor feeds informing the AI control plane.
These are the invisible machine-readable equivalents of the celebrity deep fakes grabbing headlines, but they are far more impactful. The core methodology is the same—create a corrupted imitation of what a human or AI “sees” to precipitate the attacker’s desired toxic response and objectives.
When larger quantum computers become available, these channels can be compromised with hacker-engineered data by brute force, even if they are strongly encrypted. This has the newly coined name Quantumbleed and should have a whimsical but apocalyptic CVSS severity score of 11 because it affects virtually all digital systems deployed today. There should be an AI name alluding to digital mind control à la MKUltra but paying homage to its grandfather Stuxnet—Stultra, Muxnet?
The problem is difficult, and the solutions are complex, but they are manageable if we understand that no new buzzword will fix them. The race to deploy AI with built-in quantum security must be table stakes when discussing critical systems.
These technologies are too important to our economy to ignore, but safety first. Architecting security into the core is far cheaper than remediation later. This must begin with hiring a team or assigning existing staff to develop expertise in both domains before deploying these technologies.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
