Beyond Good Intentions: Strong Algorithms, Weak Keys And The Assumptions AI Is Exposing
Denis Mandich is the Cofounder of Qrypt, a quantum cybersecurity firm, and founding member of the Quantum Economic Development Consortium.
gettyAI is revealing what enterprises never verified about the foundations of digital trust.
A cybersecurity system should not require a verdict on someone’s intentions before it can protect your business. Yet when researchers uncover unexplained choices or extraordinary mistakes, the debate often turns to motive: Was this negligence, or was it deliberate? For executives, a more urgent question comes first: Why could either explanation leave so much at risk?
Two October disclosures bring that question into focus. One reconstructs the origins of widely used cryptographic parameters, while the other exposes disastrously weak private keys inside the internet’s trust infrastructure. They are different events, but together they challenge the assumption that familiar standards and trusted institutions provide sufficient assurance. New AI tools enable top security teams to trust but verify almost everything in their cybersecurity stack.
On October 16, Anthropic reconstructed the origins of several NIST elliptic curves, mathematical structures used in cryptography. Their public parameter seeds came from a common but deprecated software generator, lrand48(), initialized with Unix timestamps. Anthropic’s latest model found the origins in under half an hour. A question that had persisted for decades suddenly had a quickly reproducible answer with no cryptanalysis expertise. Although this was not a cryptographic break, it did reveal an AI accelerated cryptanalytic technique anyone can apply. Public curve parameters are not secret keys, and their reproducibility does not make private keys predictable. The discovery also narrowed the scope for some suspected forms of manipulation, but Anthropic went further. It presented the result as grounds to close years of speculation, with no action necessary.
My concern is this confidence that an ordinary explanation can inspire. A reproducible mechanism tells us how a result arose. It does not necessarily establish why someone chose that mechanism. From an intel community perspective, always consider plausible deniability: nefarious conduct with a credible innocent explanation. An adversary who could disguise deliberate interference as a routine development shortcut would gain two advantages: the weakness itself and an explanation if someone discovered it. A familiar software function could fit either narrative.
That possibility does not establish sabotage, much less make it equally likely. It does mean apparent carelessness cannot, by itself, establish innocence, which Anthropic acknowledges as unresolved questions. For P-256, the process passed over an earlier candidate with relevant mathematical properties before it reached the published curve. The researchers discuss possible deliberate selection, but also possible unintentional causes. The evidence does not settle that choice. Executives do not need to resolve this historical uncertainty. They need systems whose security depends less on such judgments.
Days later, a separate disclosure demonstrated the consequences of misplaced confidence.
A Mozilla investigation found eight eMudhra intermediate certificates had trivially recoverable private keys. Six served TLS, the protocol that protects network connections; two served code signing, which authenticates software publishers. The keys used the P-384 curve. Yet every affected private value was an integer below 256. The mathematics had not failed, but the process that supplied its secrets had.
For clients that still trusted the affected TLS certificates, an attacker could impersonate a legitimate service. Interception would also require an appropriate network position aka man-in-the-middle, meaning, possession of these keys did not automatically decrypt all previously recorded traffic. The incident reached beyond eMudhra’s own customers because a certificate authority, an issuer of digital identity credentials, participates in a shared trust system. Other organizations inherited exposure through the certificates their software accepted.
eMudhra’s preliminary explanation described a script flaw, weaker replacement keys and confusion between key labels. The company said hardware security modules had generated the keys but its account remains subject to investigation. Nothing in the disclosed evidence establishes deliberate interference or nefarious actors. Nor must a board prove active meddling before it recognizes the architectural failure.
The response illustrates the cost of distributed dependence. Browser protections do not automatically repair every application, container or embedded device. Some clients require separate updates because they do not consume certificate revocation information. An explanation can arrive quickly while recovery depends on every affected system, which in some cases will be vast global enterprise infrastructure.
Neither event demonstrates that AI can defeat sound cryptography at will. However, the more immediate lesson is AI can help investigators reconstruct obscure processes and test assumptions that once required scarce specialist attention, frequently by state-level resources and expertise. Mozilla’s investigator also had local AI agents recover the defective certificate keys on their first attempt. Although weak keys were easy to recover without AI and easy prey for intel services, the significance is the potential to make such checks more accessible and routine to speed and scale. Defenders should exploit that capability too. Boards should ask whether their organizations actively test cryptographic assumptions or merely inherit assurances from suppliers.
Post-quantum cryptography, algorithms designed to resist quantum attacks, remains essential. It does not replace sound entropy, the unpredictability used to create keys, or controls over how applications use them. Modern cryptographic pseudorandom generators are not equivalent to lrand48(). Nor would a quantum entropy source alone prevent software from selecting the wrong key. The entire process needs scrutiny.
Executives should require evidence about entropy sources, key creation, custody and replacement. Teams should demonstrate what happens when a source fails, an administrator makes a mistake or a trusted service becomes hostile.
For high-value communications, this makes newer architectures worth serious evaluation: independently validated quantum entropy, protected delivery and endpoint key generation that reduces dependence on centralized custody. Independent inputs and separate controls can limit the damage from one compromised component, provided their independence is real and tested.
The objective is not another component labeled “trusted.” It is fewer assumptions that can fail together. The next security investment should reduce how much your business depends on everyone else having done the right thing. Complex systems break in complex ways, and sometimes that’s deliberate. Unlike the average hacker team, leadership in sophisticated services require plausible deniability that looks just like the kind of sloppy mistakes we see here.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
