On August 9, a voluntary security team claimed to have scanned approximately 150 code repositories tied to Bitcoin core projects using a suite of advanced AI models—Kimi K3, OpenAI's GPT Sol, Anthropic's Claude Fable, Opus, and Z.ai's GLM 5.2. Their reported hit rate: one critical vulnerability per researcher per hour. Over a dozen flaws across wallets, cryptographic libraries, and infrastructure components were flagged. Reports have been submitted to affected projects, though names remain under wraps.
The ledger doesn't lie. But the code that writes the ledger? That's where the truth gets tangled.
Context: The Fragile Stack Underneath Every Transaction Bitcoin's security model relies on a thin layer of battle-tested code—Bitcoin Core, libsecp256k1, HWI, and a handful of wallet implementations. These repositories are audited by humans, slowly, often after a bug has already been exploited. The voluntary team's approach flips that timeline: AI models trained on millions of lines of open-source code can pattern-match known vulnerability classes—buffer overflows, integer overflows, timing attacks, signature malleability issues—at machine speed.
The models used are not generic chatbots. They are specialized code-analysis variants: Kimi K3 for static analysis, GPT Sol for semantic understanding, Claude Fable for fuzzing logic, Opus for cryptographic verification, and GLM 5.2 for dependency mapping. The team claims each model contributes a different lens, reducing false positives through cross-validation.
But speed is a double-edged sword. In my 2020 DeFi stress test work, I simulated liquidation cascades across Compound and Aave. I learned that data patterns precede market sentiment—but also that automated detection without human context can trigger false alarms. The same applies here.
Core: The On-Chain Evidence Chain of AI-Discovered Bugs Let's examine the methodology. The team scanned 150 repos. That's a broad net—includes not just Bitcoin Core but also popular wallet libraries like BDK, rust-bitcoin, and hardware wallet firmware. The vulnerabilities reported likely fall into three buckets:
- Cryptographic implementation flaws – Incorrect scalar multiplication, non-constant-time comparisons, weak randomness in key generation. These are the most dangerous because they can lead to private key recovery. In my 2017 Chainlink audit, I traced a similar latency vulnerability in price feed aggregation that could have been exploited via flash loans. The pattern repeats: a single off-by-one in a cryptographic routine can drain millions.
- Fee logic errors – Mis-handling of fee rate calculations in transaction building. Attackers can craft transactions that appear cheap but consume disproportionate block space, or trick wallets into overpaying. This is a classic 'dust' attack vector, but with AI, attackers can now generate thousands of variants to find the exact edge case.
- Infrastructure dependency poisoning – Vulnerabilities in RPC libraries, hardware wallet communication protocols (e.g., HWI), or PSBT parsing. These allow man-in-the-middle attacks during signing. The recent Coldcard incident—where a vulnerability in the seed generation process was exploited—shows that even hardware wallets are not immune.
The team's claim of one critical bug per hour is aggressive. Based on my experience auditing oracle contracts, a thorough manual audit of a single repository can take weeks. AI accelerates discovery, but it also amplifies the noise-to-signal ratio. The real metric is not bugs found per hour, but bugs confirmed by human reviewers.
Coldcard and Boltz incidents, both occurring in the past month, underline the urgency. In the Boltz case, an attacker used AI-generated fuzzing inputs to trigger a swap contract bug, draining liquidity. The attacker's toolchain likely included similar models to those used by the security team. The asymmetry is stark: defenders and attackers now share the same AI arsenal.
Contrarian: Correlation ≠ Causation, and AI ≠ Safety Here's the blind spot most coverage misses. The security team's AI models are trained on known vulnerability patterns. They excel at finding bugs that look like bugs from the past. But the most devastating vulnerabilities—the ones that lead to consensus failures or chain reorganizations—are often novel logic errors that no training set covers.
Consider the 2018 Bitcoin inflation bug (CVE-2018-17144). It was a logic error in the validation of duplicate inputs. No AI model at the time would have flagged it because the pattern was unprecedented. The team's AI might find 100 buffer overflows, but miss the one logic bug that brings down a network.
Furthermore, the speed of AI-assisted discovery creates a pressure to patch faster than ever. Projects receiving reports now have to triage, verify, and deploy fixes within days, not weeks. This rush increases the risk of incomplete patches or new bugs introduced during hotfixes. I've seen this in institutional ETF data audits: when a custody proof discrepancy was found, the issuer had to adjust within 48 hours, leading to a second, smaller discrepancy in the revised report.
Another concern: the AI models themselves are black boxes. The team used five different models, but who audited the auditors? If a model has a hidden bias—say, it systematically misses timing attacks in Rust code because its training data was dominated by C—then entire classes of vulnerabilities remain invisible.
Code doesn't guess. But AI does. And guessing with confidence is dangerous.
Takeaway: The Next Signal to Watch Over the next two weeks, I will be monitoring two on-chain signals:
- Patch deployment frequency across Bitcoin-related repositories. A spike in commits to wallet libraries or cryptographic primitives suggests active remediation. If the commits are clustered around specific files (e.g., PSBT parsing, fee estimation), we can infer the nature of the vulnerabilities.
- Exploit attempts using AI-generated signatures. Attackers will likely weaponize the same models. Look for unusual transaction patterns—specifically, transactions that exploit edge cases in signature verification or fee calculation. On-chain data will show the fingerprints.
The voluntary team's work is valuable. But it is a starting point, not a finish line. The ledger demands human verification—always has, always will. The question is whether the industry will treat AI as a scalpel or a sledgehammer.
Data over drama. Always.