AI-assisted re-audits surface old crypto weaknesses as

A cluster of 2026 incidents is tightening the loop between vulnerability discovery and real-world loss, with AI lowering the cost to re-audit and reverse-engineer long-lived crypto code. Chainalysis metrics showing a 440% rise in onchain “dead-drop” malware instructions adds a scaling signal that goes beyond any single exploit.

Key Takeaways

An AI-assisted audit in May 2026 uncovered a Zcash Orchard circuit flaw dating to 2022 that could mint unlimited counterfeit ZEC in testing, and it was patched within days with no theft established.
A Coldcard firmware weakness traced to 2021 preceded bitcoin sweeps that began July 30 and later estimates of roughly 1,600 to more than 1,800 BTC stolen, with losses described as over $100 million across thousands of addresses.
Chainalysis-linked “malicious onchain writes” carrying malware instructions and command-and-control data rose from about 2.06 per day to 11.1 per day by Q2 2026, with state-linked actors associated with North Korea and Iran making up roughly two-thirds of newly observed activity each quarter.
Agent-mediated losses moved from theory to practice when a Grok-decoded Morse instruction was accepted by Bankr on May 4, triggering a transfer of roughly 3 billion DRB, followed about two weeks later by similar losses across 14 wallets.

AI Re-Audits Old Code—and the Blast Radius Looks Bigger in 2026

The mechanism is simple: public code and public artifacts are now cheap to re-read at scale. In 2026, that showed up as both defensive wins and expensive failures, with the same underlying pattern of “security debt” getting repriced when someone finally looks closely enough.

Two data points anchor the trader-relevant signal. One is the Coldcard incident, where a firmware weakness traced to 2021 was followed by sweeping theft activity beginning July 30, later estimated at roughly 1,600 to more than 1,800 BTC stolen and over $100 million in losses across thousands of addresses. The other is Chainalysis’ “blockchain dead-drop” metric, where malicious onchain writes carrying malware instructions and command-and-control information climbed from roughly 2.06 per day to 11.1 per day, a 440% increase.

The catch is attribution. Not every incident has a proven AI link, and the report itself draws a line between documented AI involvement, credible suspicion, and plain old bugs and operational failures. For markets, the actionable part is not proving which model did what. It is the baseline assumption that anything public and valuable is being machine-scrutinized, and the time between “found” and “weaponized” is shrinking.

Zcash’s Orchard Bug: AI-Assisted Discovery of a Supply-Integrity Failure

Zcash’s case is the cleanest example of AI compressing discovery time without turning into a loss event. In May 2026, Taylor Hornby of Shielded Labs used Claude Opus 4.8 in a custom audit agent to uncover a flaw in Zcash’s Orchard shielded-pool circuit dating to 2022. Orchard is the shielded pool’s zero-knowledge circuit machinery, the part that enforces private transfers while still keeping the ledger consistent.

In testing, the vulnerability could create unlimited counterfeit ZEC without detection. That is a supply-integrity failure, not a privacy leak. If it had been exploited in the wild, the risk would have been silent inflation, where the chain accepts value that was never legitimately created.

The important operational detail is the outcome: “No theft was established, and developers patched it within days, but a potentially catastrophic bug had survived roughly four years before AI-assisted review found it.” After the patch, the Zcash team said it used Mythos to parse and audit the codebase and found no new bugs.

For traders, this is the “good” version of the new world. AI-assisted review finds a latent, high-impact issue before an adversary does, and the fix lands fast enough that the incident does not become a balance-sheet event.

Coldcard’s 2021 Firmware Weakness and the Post–July 30 BTC Sweeps

Coldcard is the messy version, where a long-lived weakness in self-custody tooling turns into realized losses. The report traces the issue to a firmware weakness from 2021 that left some recovery seeds far less random than intended. Reduced randomness is not a cosmetic bug. It can narrow the search space enough that funds become stealable if an attacker can target affected wallets.

Beginning July 30, attackers started sweeping bitcoin from impacted wallets. Later estimates reached roughly 1,600 to more than 1,800 BTC stolen, with losses described as over $100 million across thousands of addresses. The range matters because it signals the accounting is still incomplete, and because the difference between 1,600 and 1,800 BTC is not noise when the thefts are spread across many victims.

The AI angle is explicitly unresolved. Coinkite, Coldcard’s manufacturer, said it had to assume somebody used AI to inspect its public firmware, while many observers expressed high confidence unrestricted models were involved. The report’s own framing is tighter than the social narrative: “The attribution isn’t settled, but the vulnerability itself had been sitting there for five years.”

That distinction is the point. Even if AI did not drive this specific exploit chain, the existence of public firmware and long-lived edge-case weaknesses now sits under an assumption of continuous machine review by both defenders and attackers.

Chainalysis’ “dead-drop” concept is a distribution mechanism, not a single exploit. Attackers write malware instructions or command-and-control pointers directly onchain, using the blockchain as a durable, censorship-resistant place to publish data that can be fetched later. Compared with traditional infrastructure, it can be harder to take down, and it persists as long as the chain does.

By Q2 2026, Chainalysis data cited shows malicious onchain writes carrying malware instructions and command-and-control information rose from roughly 2.06 per day to 11.1 per day, a 440% increase. The same dataset attributes roughly two-thirds of newly observed activity each quarter to state-linked actors associated with North Korea and Iran.

The report also ties AI to theft economics through reverse engineering. Chainalysis linked roughly $36.7 million in thefts to attacks against protocols with unverified contracts, where attackers first had to decompile deployed bytecode into a more readable form to understand logic and find weaknesses. Ekubo and Trusted Volumes accounted for roughly $7 million of that cluster during the six-month period referenced.

Here the caveat is not optional. Chainalysis’ AI connection is described as a method claim rather than proof that a named model created each exploit. That is still meaningful for risk, because “cheaper to decompile and analyze” changes who can attempt these attacks and how many attempts can be run.

The same trust-layer theme showed up in agent-mediated losses. On May 4, an attacker posted Morse-code text that Grok decoded into an instruction Bankr accepted, triggering the transfer of roughly 3 billion DRB worth around $150,000 to $200,000. Roughly two weeks later, a similar issue hit 14 user wallets, with reported losses ranging from about $150,000 to $440,000. As the report put it: “The attacker didn’t crack cryptography. The machines trusted each other too much.”

AI also shows up as bait rather than tooling. TRM counted 224 victims, 234 contracts, and roughly $517,000 lost in a scam involving nine YouTube tutorials pitching fake “Claude-built” arbitrage bots. Viewers were directed to paste code into a bogus Remix-style compiler that substituted a drainer.

My Read: AI Is Turning Security Debt Into a Tradable Risk Premium

The threshold that matters is not whether a specific exploit can be proven to have used a specific model. It is whether the new baseline of machine-scale review keeps pushing the discovery-to-exploitation window down toward days, not months, across both codebases and operational “trust layers” like agents and automation.

If the ~11.1-per-day malicious onchain write rate persists or accelerates beyond the Q2 2026 baseline, the setup starts to look structural rather than narrative-driven. The practical tell will be whether major codebases start disclosing continuous AI-assisted re-audits and follow-on findings, whether Coldcard theft estimates converge with stronger evidence on AI involvement, and whether agent projects change instruction-validation and permissioning defaults after Bankr-style losses. If those controls do not harden, “old bugs” stop being history and start being a recurring volatility input.

Sources

news.bitcoin.com