7f4e681e-0

ConfirmedSecurity741 words4 min decode

StarkWare CEO Ben-Sasson: AI Finds Bugs but Cannot Fix Them

StarkWare CEO Eli Ben-Sasson questioned on November 18 why AI reliably detects security flaws but cannot patch them, weeks after warning AI attacks may precede quantum threats.

Eli Ben-Sasson questions why AI can find bugs but not fix them
WitnessEli Ben-Sasson questions why AI can find bugs but not fix themStockSnap / pixabay

Outputs

  1. Eli Ben-Sasson posted on X on November 18, 2026 asking why AI can find vulnerabilities but not fix them.

  2. OpenAI reported its AI pipeline discovered and patched hundreds of Linux kernel vulnerabilities.

  3. On November 5, 2026, Ben-Sasson warned AI-driven attacks could emerge before quantum threats.

  4. Ben-Sasson advocates hash-based cryptography and ZK-STARKs as resilient alternatives.

  5. Ben-Sasson co-founded Zcash and leads StarkWare, built on ZK-STARK technology.

StarkWare CEO Eli Ben-Sasson asked on November 18, 2026 why artificial intelligence can uncover hidden security vulnerabilities in software but cannot reliably repair them. His post on X did not name a protocol, company, or incident, and reads as expert commentary on a structural gap in how AI systems handle security work.

Ben-Sasson leads StarkWare and co-founded Zcash. He is best known for his contributions to zero-knowledge proofs, which let one party prove a statement is true without revealing the underlying data.

Why does detection outpace repair?

The asymmetry Ben-Sasson highlighted has a structural logic, according to his framing. Finding bugs rewards breadth: a model can scan everything and flag anything suspicious. Fixing bugs rewards depth: a repair requires understanding the whole system well enough that a change in one place does not cause trouble somewhere else.

The remediation side has not kept pace with detection. Spotting a broken lock is one skill. Installing a new one without leaving the back door open is another.

That distinction carries real operational weight, because a bad patch can be worse than no patch at all. A known bug is at least a known, trackable risk. A fix that quietly introduces a fresh flaw creates a risk nobody is monitoring — a failure mode that is harder to price, audit, or defend against than the original defect.

What did OpenAI's Linux kernel work show?

Ben-Sasson's remarks arrive alongside discussion of OpenAI's recent security results. OpenAI reported that its AI pipeline discovered and patched hundreds of vulnerabilities in the Linux kernel.

The concern sits in the second half of that result. AI-generated patches have been observed to sometimes introduce exploitable vulnerabilities of their own, which is precisely the failure mode Ben-Sasson's question points at.

Read together, the OpenAI data and Ben-Sasson's post suggest a split scorecard for AI in security. The headlines about AI finding hundreds of bugs cover only the detection half. Remediation remains a separate discipline with a separate — and so far weaker — track record.

Is this part of a broader warning?

This is not the first time this month that Ben-Sasson has flagged AI as a security concern. On November 5, 2026, he warned that AI-driven attacks could potentially emerge ahead of quantum threats.

His earlier message was that preparation should not wait. He emphasized moving toward more resilient cryptographic frameworks, including hash-based systems and ZK-STARKs.

The two technologies differ in their assumptions. ZK-STARKs are the proof system at the core of StarkWare's product line and rely on hash functions rather than elliptic-curve assumptions. Hash-based approaches similarly lean on hash functions instead of the number-theory problems that quantum computers are expected to threaten.

Read as a sequence, the two posts sketch a two-sided problem. AI may sharpen attackers before quantum computers arrive, and AI is not yet a dependable defender on the repair side. Attack capability and defensive capability are improving at different speeds, and the gap between them is where exposure accumulates.

What does this mean for crypto teams?

For crypto projects, the implications are direct. Blockchains, DeFi protocols, and identity systems run on cryptography and on code that frequently controls real money, which raises the cost of both unpatched bugs and botched patches.

The practical takeaway for development teams follows from the detection-versus-repair split. AI can serve as a useful tool for surfacing problems, but its proposed fixes still require careful human review before deployment anywhere near production systems. Treating discovery and remediation as separately judged capabilities, rather than a single "AI security" function, is the operational discipline Ben-Sasson's post implies.

His advocacy also signals where security spending could flow. Sectors that depend heavily on cryptography — DeFi and identity among them — may see increased interest in hash-based systems and ZK-STARKs if the emphasis on post-quantum resilience keeps growing.

Ben-Sasson leads a company built on one of the technologies he is recommending, so his advocacy carries a commercial dimension alongside the technical argument. Readers can weigh both.

The immediate signal for security teams is procedural: discovery and remediation should be measured, staffed, and audited as distinct functions. Until AI-generated patches stop introducing their own vulnerabilities, the human review layer between a detected flaw and a deployed fix remains the control point that determines whether a known risk stays known.

via Crypto Briefing (Source)

More from Tom Whitfield

Tom Whitfield

Show full bio

News editor covering technology and product news in crypto and web3 industry at Mempool Brief.

1,440 articles