Cryptographers ask AI labs to test encryption
Oct 14, 2026Wednesday, October 14, 2026Summary from 5 contributing podcasts.
- No show reported an AI break of Bitcoin’s signature cryptography; the new push is for labs to test it directly.
- Dan Boneh says public results could clarify whether core mathematical assumptions hold or need replacing.
- Experts warn that hurried wallet moves can create immediate losses, while software audits and crypto agility offer defenses.
The latest development in the AI-and-encryption debate is a call for evidence, not proof of a break. On Oct 13, 2026, The a16z Show reported Stanford cryptographer Dan Boneh’s appeal for frontier AI labs to test standard cryptographic assumptions and share the results. OpenAI’s mathematical papers, the show noted, did not address cryptography.
That gap matters because systems including Bitcoin’s signatures rely on mathematical problems that have resisted attack, not on a proof that they are impossible to solve. On Oct 8, 2026, Bitcoin And described host David Bennett’s skepticism about panic over AI and elliptic-curve cryptography. On Oct 9, 2026, Bankless reported Justin Drake’s warning that AI could accelerate a break, alongside objections that the danger remained unproven.
Boneh’s proposed test could turn that argument into a more useful signal. As The a16z Show relayed it, he said substantial compute could compress years of human cryptanalysis into a much shorter test: a successful attack would warn developers, while failure would add evidence for the assumptions’ security. That would not prove a system safe forever, but it could replace speculation with targeted results.
The show also made the case that AI can strengthen defenses. A16z crypto general partner Eddie Lazarin argued that cheaper AI tools can repeatedly scan software for finite, discoverable vulnerabilities. Those audits address bugs in implementations - not a break in the mathematics itself - and offer a practical security gain even while the cryptographic question remains open.
For users, the distinction changes what a sensible response looks like. The a16z Show said major internet systems have ways to switch cryptographic methods and cited hybrid encryption in Signal and iMessage. But it did not establish that Bitcoin can make such a change without difficulty. A16z research partner Justin Thaler warned that rushed wallet migrations can cause losses through human error. Until labs publish test results, the risk of a mathematical shortcut remains unknown; panic transfers are not a substitute for preparation.
Source Intelligence
- Deep dive into what was said in the episodes
Can AI Break Cryptography? What the Experts Actually Think • Oct 13
- OpenAI released 700 research papers demonstrating mathematical progress, but none addressed theoretical or applied cryptography. Dan Boneh urges frontier labs to publicly share model performance when attempting to break standardized cryptographic assumptions.
- Security researcher Justin Drake urged institutions to enter bunker mode over fears AI could break elliptic curve cryptography quickly. Justin Thaler warns that hasty asset migrations historically trigger more financial loss through human error than direct cryptographic hacks.
- Public key cryptography relies on unproven mathematical complexity assumptions tied to the P versus NP problem. Dan Boneh notes that AI models failing to break these assumptions provides stronger empirical evidence for their long term security.
- Validating new cryptographic assumptions historically required decade long academic scrutiny before deployment. Eddie Lazarin claims labs can now gain equivalent confidence in weeks by directing high capability AI models to aggressively attempt targeted breaks.
- Intelligence token costs dropped up to 99 percent over recent model generations, enabling affordable continuous codebase auditing. Eddie Lazarin and Dan Boneh predict cheap agentic AI tools will systematically eliminate software vulnerabilities, net-improving global security posture over time.
- Internet infrastructure utilizes crypto agility to switch primitives rapidly when vulnerabilities emerge. Dan Boneh highlights that major web traffic protocols already run hybrid encryption standards like X25519 combined with ML-KEM-7, preserving protection if one algorithm fails.
Also discussed on this episode: (3)
Safety (3)
- Breaking elliptic curve discrete log would collapse global digital infrastructure across banking, government communications, and healthcare. Eddie Lazarin notes that major institutions already employ standard key management protocols to mitigate private key exposure on public blockchains.
- Blockchain developers debate adopting hash-based signatures over lattice-based signatures for post-quantum security. Dan Boneh explains hash schemes avoid adding new mathematical assumptions, while lattice systems provide rich algebraic capabilities like threshold signatures and non-hardened key derivation.
- Enforcing hash-only cryptography in blockchain networks requires complex SNARK recursion to maintain efficiency. Justin Thaler argues this approach drastically expands software attack surfaces and relies on mathematical security properties that cryptographers do not yet fully understand.
The AI Daily Brief: Artificial Intelligence News and Analysis
Nathaniel Whittemore
What Happens When AI Solves Your Life’s Work • Oct 9
- OpenAI published 372 novel mathematical results and 722 supporting papers generated by an unreleased frontier model, fully solving 90 of ProofAtlas's top 500 open problems. The model averaged three hours of ChatGPT Pro compute per result.
- Bitcoin researcher Justin Drake highlighted a notable absence of cryptographic breakthroughs in OpenAI's publication, suggesting potential government intervention. Computer scientist Scott Aaronson confirmed AI laboratories are privately testing whether internal frontier models can break major cryptographic primitives.
Also discussed on this episode: (7)
Startups (1)
- Financial reports revealed OpenAI told investors its annualized revenue reached $50 billion at the end of September, falling short of a $68 billion whisper number derived from gross revenue comparisons with Anthropic. Semiconductor stocks dropped sharply following the report.
Enterprise (2)
- A survey by The Information found 66% of subscribers work at organizations building AI applications, with 35% reporting multi-fold returns on AI spending. AI adoption had no impact on hiring practices for nearly half of respondents.
- OpenAI regained top market share among tech enterprise subscribers ahead of Anthropic and Google. Open-weight models achieved a 72% adoption rate, though only 21% of respondents deploy Chinese open-weight models due to organizational hesitancy.
Safety (1)
- Anthropic updated its terms of service to prohibit abusive behavior toward Claude, a move Aaron Levie defended as essential for preventing negative training data feedback loops. Anthropic also launched Claude Dashboards and Claude Motion for animated data visualization.
Reasoning (2)
- Rutgers professor Alex Konturovich noted that a human producing the quasi-Riemann hypothesis result would instantly earn a Fields Medal. University of Toronto mathematician Darren Litt framed the release as a structural pivot toward human verification of machine-generated proofs.
- Mathematician Francesco Maggi warned that unexamined AI proofs remain dead letters if human culture fails to absorb and contextualize them. Physicist Steve Hsu cautioned that overwhelming research volumes could render theoretical scientific disciplines uninterpretable to human researchers.
Models (1)
- Wharton professor Ethan Mollick reported AI models are producing top-journal quality research in economic sociology. Computer scientist Pedro Domingos and Google researcher Payman Milanfar explained that low-entropy domains automate quickly, whereas complex fields like medicine and law remain resistant.
Keeping AI or SI open, Will math discoveries break ECDSA, Pathways for agentic payments • Oct 9
- Justin Drake urged users to migrate Bitcoin to fresh addresses over elliptic curve security concerns. DK counters that moving funds offers zero protection if ECDSA is broken, as spending transactions inevitably expose public keys to the network.
- Steve explains that signing software with implementation bugs or improper nonce generation can leak private key material to blockchain observers. Repeated transaction signatures under buggy software allow external attackers to deduce private keys without breaking underlying cryptography.
- DK argues artificial superintelligence can compress 20-year cryptographic testing periods into two days by simulating and breaking novel algorithms. This acceleration could allow Bitcoin to safely adopt complex post-quantum schemes like lattice-based cryptography faster.
- Hash-based signature schemes lack support for Hierarchical Deterministic wallets, Schnorr signatures, and standard key management features. Adopting them forces exchanges and individuals to store and back up millions of distinct private keys independently.
- DK warns that a sudden break in ECDSA would expose up to 9 million Bitcoin to immediate theft. Such an attack would destroy systemic confidence and require controversial social consensus to halt and roll back the blockchain.
- Anthropic introduced a security scanning initiative for open-source codebases using proprietary closed models. DK highlights that Anthropic's mandatory 90-day disclosure rule conflicts with Bitcoin Core's policy of withholding critical vulnerability details for years to protect unpatched nodes.
Also discussed on this episode: (5)
Open Source (1)
- DK proposes adopting Open Super Intelligence or Open Intelligence as standard terms for open-weight models. Steve points out that OSI creates a brand conflict with open-source intelligence used in government security sectors.
Models (1)
- DK explains that Nvidia cannot release Nematron with fully transparent training datasets due to commercial data licensing agreements. Open-weight models remain the most practical compromise for widespread deployment despite lacking cryptographically verifiable inputs.
Custody (1)
- Fraudulent hardware distributed through a Southeast Asian reseller enabled attackers to steal $86 million from users who deployed compromised devices. Steve emphasizes that hardware attestation protocols like Bitkey's remote verification are required to protect users against reseller tampering.
Agents (2)
- Max demonstrates setting up autonomous Grokbot agents with self-custodial Bitcoin wallets to purchase domains and Lightning services without human intervention. Max argues that combining autonomous software with native digital payments enables agents to run real businesses end to end.
- DK uses deterministic software control planes to govern probabilistic large language models before executing sensitive actions like sending emails. DK advocates for hybrid architectures where deterministic rules explicitly bound model capabilities to prevent catastrophic autonomous failures.
ROLLUP: AI Threatens Crypto Security | Tom Lee Stops Buying ETH | Ethereum L2s Shut Down | Zcash ETFs • Oct 9
- Justin Drake urged crypto users to initiate a controlled migration to fresh addresses, citing rapid AI mathematical progress that could break ECDSA private key cryptography. OpenAI solving 377 complex math problems indicates superintelligence might compromise public key cryptography far sooner than quantum computing.
- Cryptographer Matthew Green warned AI advancement threatens standard public key encryption across global banking and communication infrastructure. Conversely, Coinbase cryptography leaders and Mert Mumtaz dismissed bunker recommendations as performative, arguing fundamental mathematical hardness assumptions remain unbroken.
Also discussed on this episode: (9)
Macro (1)
- A macro combination of elevated Treasury yields and oil breaking $100 per barrel continues to pressure crypto markets. High yield environments accelerate government debt service burdens, driving long term fiat debasement narratives even if short term liquidity tightens.
Adoption (1)
- Robinhood completed its inaugural corporate treasury purchase of crypto by acquiring $25 million in Bitcoin. The firm explicitly cited direct alignment with the broader cryptocurrency ecosystem for the balance sheet addition.
Markets (1)
- Tom Lee announced Bitmine is capping its Ethereum acquisition strategy after buying nearly 6 million ETH to reach its 5% supply target. Bitmine will direct its $643 million cash reserves toward stock buybacks, staking infrastructure, and strategic investments.
Protocol (2)
- The Lean Ethereum research roadmap is prioritizing hash-only signature schemes over lattice-based cryptography to mitigate emerging AI and quantum vectors. Vitalik Buterin confirmed that hash-based primitives offer superior resistance against rapid mathematical optimization.
- Ethereum L2s Blast and Abstract announced shutdowns due to unsustainable standalone business models, erasing Blast's previous peak TVL. Meanwhile, Starknet revealed plans to transition into an independent quantum-resistant L1 network target by 2027.
Regulation (2)
- SEC Commissioner Hester Peirce concluded her term at the regulatory agency. Her early 2020 Token Safe Harbor framework laid the foundation for the innovation exemption policy adopted under Chairman Paul Atkins.
- FinCEN formally withdrew proposed 2020 regulations that required exchanges to report non-custodial wallet transactions exceeding $3,000. The agency also abandoned proposed rules that would have universally blacklisted assets interacting with privacy mixers.
ETFs (1)
- Grayscale's Zcash ETF expanded past $1 billion in assets under management relative to the protocol's total market cap. Concurrently, the Winklevoss twins submitted filings for a spot Zcash ETF under the ticker WINK.
Agents (1)
- Token 2049 in Singapore brought together up to 20,000 participants, with discussions centering heavily on AI agents, privacy tools, and institution-led expansion. Hasib Qureshi noted that early expectations for autonomous AI agent payment rails on-chain have yielded to agent usage of legacy credit card rails.
AI Scares Quantum | Bitcoin News • Oct 8
- Ethereum researcher Justin Drake warned users to enter bunker mode, claiming AI could break wallet elliptic curve mathematics within months. David Bennett rejects the panic, noting the Cold Card hack stemmed from faulty random number generators, not cracked cryptography.
Also discussed on this episode: (7)
Custody (1)
- The US government transferred 9,261 Bitcoin worth $770 million to Coinbase Prime, pulling from Bitfinex exploits and Binance seizures. This movement conflicts with a March 2025 executive order directing seized crypto into a strategic reserve.
Regulation (1)
- CFTC Chair Mike Selig claims a new regulatory framework targeting leveraged crypto exchanges will prevent future FTX-style collapses. David Bennett argues regulatory regimes cannot stop human malice, citing the spectacular failures of Lehman Brothers and Enron.
Europe (1)
- European Parliament lawmakers passed a non-binding resolution pushing crypto into the European Commission's upcoming anti-corruption strategy slated for 2026. David Bennett argues the policy is born from political fear of Bitcoin rather than actual systemic corruption.
Stablecoins (1)
- Securitize launched 1:1 backed tokenized US stocks on Solana using Convertible Entitlement Tokens settled in USDC. The on-chain value of tokenized stocks has surpassed $3 billion, with Securitize accounting for the bulk of recent market growth.
Safety (2)
- Vitalik Buterin agreed AI advances threaten lattice-based cryptography, but urged holders against rushed migrations. Buterin revealed he has lost more personal wealth to botched migrations and transaction errors than to actual smart contract hacks.
- David Bennett suggests that rising AI-driven cyber threats targeting critical infrastructure and databases may eventually force society back toward notarized paper records to secure physical wealth, deeds, and asset ownership.
Agents (1)
- Hackers deployed automated AI sub-agents to breach South Korea's Yoido Full Gospel Church database, exposing the records of 850,000 members. Oasis Security identified machine-written logs detailing the coordinated multi-agent attack.