AI Agents Are Finding Decades-Old Software Bugs. Should You Worry?
AI models are now discovering long-hidden vulnerabilities in banking, crypto, and open-source code faster than humans ever could. Here's what changed.
What’s actually happening
AI models are now capable of scanning massive codebases and cryptographic systems and finding flaws that sat undetected for years, sometimes decades. This isn’t speculative. A recent theft of Bitcoin from hardware wallets, reportedly worth more than $100 million across thousands of addresses, traced back to a key-generation bug that had existed in the wallet’s firmware for roughly five years. Security researchers and AI labs, including Anthropic, have separately shown that advanced models can break cryptographic schemes that survived years of expert human review. The pattern is the same in both cases: code that looked solid to humans wasn’t solid at all, and AI found the gap fast.
TL;DR
- A hardware wallet bug in Coin Kite’s Cold Card device generated private keys from a much smaller, weaker pool than intended, letting an attacker precompute possible keys and drain wallets once matched to funded public addresses.
- The theft moved fast: attackers reportedly drained around 1,200 wallets in about 41 minutes, targeting the largest balances first, with total losses estimated above $100 million across thousands of addresses.
- The bug had existed for roughly five years before anyone found it, but was reportedly discovered and exploited within about two weeks of the open-source release of a highly capable model believed to be a distillation of Anthropic’s more restricted “Fable” model.
- Anthropic has publicly described using its models to find new attacks against post-quantum cryptography candidates like Hawk, breaking assumptions that had survived two years of expert scrutiny in roughly 60 hours of model run time.
- Major financial institutions are reportedly already using AI-assisted code review internally (referenced in the source as Project Glass Wing) precisely because they expect attackers to be doing the same thing, but many smaller banks and most crypto platforms don’t have equivalent access.
- The core shift is that open-source code and public firmware can no longer be assumed safe by default: if it’s ever been public, security researchers now treat it as something actively being crawled by AI systems, for both defensive and offensive purposes.
- Recovery from AI-discovered crypto exploits is often effectively impossible once funds move, because decentralized systems like Bitcoin have no central authority to reverse transactions.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
How did AI find a bug that survived five years of scrutiny?
The Cold Card wallet vulnerability wasn’t a design flaw in Bitcoin’s cryptography. Bitcoin’s private keys are supposed to be drawn from an enormous, effectively unguessable pool of random numbers, paired with a public key that anyone can see but can’t reverse-engineer. The problem was in implementation: the wallet’s firmware generated keys from a far smaller pool of possible values than it should have. That’s a narrow, specific coding mistake, the kind of thing that’s easy to miss during manual review but trivial for a system built to systematically enumerate possibilities and cross-reference them against real-world data.
Once that flaw was found, exploiting it didn’t require anything exotic. An attacker could generate every plausible weak key offline, derive the matching public addresses, then check those addresses against the blockchain to see which ones held funds. From there it’s a matter of scripting withdrawals, starting with the largest balances. The hard part was finding the bug in the first place, and that’s exactly the kind of pattern-matching-across-huge-search-spaces task that AI models handle differently than human auditors, who tend to review code for logic errors rather than statistically test assumptions about randomness.
Why is this bigger than one crypto hack?
The Coin Kite CEO, Rodolfo Novak, reportedly described this as a “sober reality of a new AI paradigm,” where AI-assisted code review can find latent bugs faster than even experienced security teams. His broader point matters more than the specific incident: any firmware or code that has ever been public should now be treated as something under active review by both attackers and defenders, using AI tools capable of processing it at a scale no human team can match.
That logic doesn’t stop at crypto wallets. Banks, payment processors, and everyday software all rest on layers of third-party code, open-source libraries, and legacy systems that were never expected to face this kind of scrutiny. A bank can secure its own codebase perfectly and still be exposed through some dependency it doesn’t fully control. Anthropic has reportedly demonstrated this directly to lawmakers, showing an AI model identifying a way to drain funds from a banking system in a controlled test, then patching the same vulnerability. The demo was reportedly enough to prompt emergency meetings among major financial regulators and institutions.
Is this the same AI doing the attacking and the defending?
Largely, yes, and that’s part of what makes this moment strange. The same capability that lets an AI model find a cryptographic weakness for a company trying to secure its own product is what lets an attacker find that same weakness first if they get access to a comparably capable model. The gap between “AI as security tool” and “AI as attack tool” isn’t a difference in the technology, it’s a difference in who’s holding it and what guardrails, if any, are attached.
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
This is why the timeline around the Cold Card hack matters. A highly capable but unreleased model (referred to in discussions of this trend as “Mythos”) was reportedly followed by a more guardrailed public version. Later, an open-source model believed by some in the AI community to be partly distilled from that guardrailed version was released without the same restrictions. The Bitcoin theft reportedly happened within about two weeks of that open release. None of this proves direct causation, but the sequence illustrates the risk: capability that was contained inside a lab eventually finds its way into tools with far fewer restrictions, and the gap between “a lab can find this” and “anyone can find this” is shrinking.
What does this mean for cryptography itself?
Beyond finding bugs in existing systems, AI models are also being used to probe the mathematical foundations of encryption itself. Anthropic has reportedly published research showing one of its models finding a new attack approach against Hawk, a post-quantum digital signature candidate that had been reviewed by cryptography experts for two years without anyone identifying the flaw. Reports describe the model cutting the effective strength of that encryption scheme in half after roughly 60 hours of analysis.
Hawk was a candidate for future encryption, not something already deployed, so the immediate practical risk was limited. But the underlying capability is the concerning part: models aren’t inventing entirely new mathematics, they’re recombining known techniques in ways human researchers hadn’t tried, and doing it fast enough to outpace years of expert peer review. As the industry moves toward post-quantum encryption standards generally, this adds pressure to test new schemes against AI-assisted attacks, not just traditional human cryptanalysis.
Is there anything ordinary users can actually do?
There’s no way for an individual to audit the firmware in their hardware wallet or the backend systems of their bank. The realistic response is basic risk reduction: keep less value concentrated in any single wallet or account where possible, pay attention to security advisories from hardware and software vendors, and treat “this has been reviewed for years” as weaker reassurance than it used to be. Institutions with access to AI-assisted security review, larger banks running internal red-teaming programs, are ahead of this curve. Smaller platforms, many crypto projects, and most consumer software are not, which means the vulnerability gap is likely to keep surfacing unevenly across the industry rather than all at once.
Frequently Asked Questions
What caused the recent Bitcoin hardware wallet hack?
A firmware bug in Coin Kite’s Cold Card wallet generated private keys from a much smaller pool of possible values than intended, instead of the full random range Bitcoin’s security model requires. This let an attacker precompute the weaker keys, match them to funded public addresses, and drain wallets.
How much Bitcoin was stolen?
Reports place losses at more than $100 million across close to 10,000 affected addresses, with roughly 1,200 wallets drained in about 41 minutes during the initial attack window.
Can AI really find vulnerabilities humans missed for years?
Yes. Both the wallet bug (undetected for around five years) and cryptographic weaknesses in a post-quantum encryption candidate (undetected after two years of expert review) were reportedly identified by AI-assisted analysis after human review failed to catch them.
Are banks using AI to find these bugs too?
Some are. Larger financial institutions reportedly use AI-assisted code review internally, referenced in discussions of this trend as Project Glass Wing, alongside tools from major AI labs. Smaller banks, many European institutions, and most crypto platforms don’t yet have comparable access.
Can stolen crypto from these hacks be recovered?
Usually not. Once funds move on a decentralized network like Bitcoin, there’s no central authority that can reverse the transaction. Recovery generally depends on the attacker making an operational mistake that allows investigators to trace and identify them.