Artificial intelligence is reshaping how developers and attackers find security flaws in blockchain systems, but recent high-profile incidents reveal a critical limitation: AI tools can identify individual coding errors while completely missing dangerous interactions between components that create real-world exploits. As the crypto industry faces record-breaking security breaches and a wave of project shutdowns, the question of whether AI-powered audits can replace human security review has become urgent.

What Happened With the Coldcard Hardware Wallet Flaw?

The Coldcard hardware wallet incident brought this tension into sharp focus. Coldcard devices, which store cryptocurrency offline and are considered among the safest ways to hold digital assets, contained firmware problems that weakened the randomness used to generate wallet seed phrases. These seed phrases are supposed to contain enough entropy, or unpredictability, to make them practically impossible to guess. Some affected devices generated seeds with far less randomness than intended, making them vulnerable to theft.

Coinkite, the company behind Coldcard, identified vulnerable firmware across several models. The company warned that simply updating the firmware would not secure an existing weak seed; users needed to create a new seed and transfer their funds to safety. Block's security researchers later published a technical analysis tracing the problem to predictable random-number fallbacks and weak reseeding behavior in older firmware. Initial reports linked the flaw to the theft of 594 Bitcoin, but later estimates placed losses close to $90 million, with subsequent reports suggesting the total exceeded $100 million as investigators identified more affected addresses.

What makes this case particularly relevant to AI security is what happened next. Coinkite noted that Coldcard's source code had always been publicly available, leading the company to assume that someone may have used AI to review older firmware and discover the weakness. However, that statement remains an assumption; Coinkite has not released evidence proving attackers used AI. More troubling, Coinkite's own AI-assisted code review failed to identify the flaw. According to a Cybernews report, later tests with advanced AI models also failed to detect the full problem.

Why Are AI Security Tools Missing Critical Vulnerabilities?

The Coldcard case exposes a fundamental weakness in current AI crypto security tools. A model may find individual coding errors while missing dangerous interactions between components. This gap between finding isolated bugs and understanding how those bugs combine to create exploitable vulnerabilities represents a serious blind spot in AI-assisted auditing.

The Bitcoin Red Team, a volunteer initiative launched in response to the Coldcard incident, offers a window into both the power and limitations of AI-assisted security review. Developers including Calle and AnchorWatch chief executive Rob Hamilton coordinated an effort where about 16 specialists reviewed more than 390 repositories in 27.5 hours using over $40,000 in AI computing resources. The team filed 4,962 findings, including 85 critical and 635 high-severity reports, which they privately disclosed to maintainers before discussing them publicly.

Those numbers demonstrate the speed of AI-assisted analysis. However, findings do not automatically represent confirmed vulnerabilities. Human specialists must reproduce each issue, remove false positives, and assess its real impact. The team plans to release its security-testing framework as open-source software, which could help other projects conduct similar reviews.

How Should Crypto Teams Balance AI and Human Security Review?

Security experts agree that AI-assisted auditing is becoming a useful defensive layer, allowing teams to review more code and test more scenarios in less time. However, the consensus is clear: AI cannot replace sound engineering practices. Teams should combine AI reviews with multiple layers of protection:

• Independent Human Audits: Professional security firms should conduct manual reviews to catch component interactions and logic flaws that AI misses.
• Reproducible Security Tests: Automated testing frameworks should verify that vulnerabilities identified by AI can actually be exploited in real conditions.
• Strict Access Controls: Limit who can approve code changes and deploy updates to production systems.
• Multisignature Wallet Protection: Require multiple approvals before moving funds, reducing the impact of a single compromised key.
• Formal Verification for Critical Components: Use mathematical proofs to verify that the most sensitive code behaves as intended.
• Manual Approval for High-Value Agent Transactions: Require human sign-off before AI agents execute large financial transactions.
• Responsible Vulnerability Disclosure: Establish clear processes for reporting and fixing security issues before they become public.

The central lesson is clear. AI can strengthen crypto security, but it cannot replace sound engineering. Attackers and defenders now have faster tools, and the safest projects will use those tools while maintaining strong human oversight.

What New Risks Do AI Agents Create in Crypto?

Beyond code review, AI agents introduce another area of concern. An agent may receive permission to access a wallet, sign transactions, or interact with smart contracts. These capabilities turn a manipulated AI response into a potential financial action. Attackers could use prompt injection, a technique where malicious text embedded in websites, messages, or documents influences an AI system's behavior, to hijack an agent and cause harmful actions.

A compromised agent might disclose sensitive information, call an unauthorized tool, or approve an unintended transaction. The US National Institute of Standards and Technology (NIST) has warned that indirect prompt injection can hijack agents and cause harmful actions. NIST's AI-agent security research recommends stronger testing and controls for systems that process untrusted information. The OWASP prompt-injection guidance also identifies unauthorized actions and data disclosure as major risks.

The broader context makes this urgency clear. The crypto industry is experiencing record-breaking security breaches. A Blockaid report estimates that $1.1 billion was lost to on-chain exploits in the first half of 2026 alone, more than all of 2025 combined. April 2026 was the most-hacked month in crypto history by number of attacks, with two attacks accounting for the majority of losses: a $293 million exploit of Kelp DAO on April 18 and a $285 million theft from Drift Protocol on April 1, in which North Korean-affiliated hackers spent six months socially engineering their way into the Solana-based exchange.

Alongside these exploits, over 100 crypto projects have shut down, filed for bankruptcy, or gone permanently dark in 2026, according to data from RootData. The exits span every layer of the industry, including exchanges, wallets, DeFi lending protocols, NFT marketplaces, and layer-1 blockchains. With venture capital rescue funds drying up, single hacks are forcing immediate protocol bankruptcies while leaving abandoned, unmaintained "zombie contracts" running on-chain.

"There isn't a venture-backed business in governance tooling for decentralized protocols, at least not yet," stated Dennison Bertram, co-founder of Tally, a DAO tooling platform that powered governance for over 500 protocols including Uniswap, Arbitrum, and ENS before shutting down.

Dennison Bertram, Co-founder of Tally

In this environment, the pressure to deploy code quickly and the temptation to rely solely on AI-assisted audits is intense. Yet the evidence from Coldcard, the Bitcoin Red Team, and the wave of exploits in 2026 all point to the same conclusion: security is not a feature that can be automated away. It requires sustained human attention, rigorous testing, and a culture that treats vulnerability disclosure as a feature, not a liability.