The Debate Over AI-Accelerated Vulnerability Discovery

Episode 1 · Cryptographer Matthew Green: Assume AI keeps improving at this year's pace, security will split into fast and slow defenders (2026-10-12, 7 posts)

On October 12, Johns Hopkins cryptographer Matthew Green laid out his view on AI progress and information security: assume models keep getting stronger at roughly the same pace as since the start of this year, treat that as the extrapolation baseline; if so, we are already in fairly new territory, and by next October it will be entirely new ground. He stressed that those who believe "nothing unexpected will happen" simply haven't noticed the many surprises this year—the only certainty is that things will be surprising.

Confirmed

• Green stated clearly that despite the hype, AI breaking cryptography is the least of his information-security worries.
• What truly concerns him is how badly network infrastructure is prepared for "what may arrive next year," and how surprisingly low the industry's estimates of upcoming attack capabilities are.
• Security expert Dino Dai Zovi responded that the real problem is not AI breaking crypto, but that most organizations don't know how to prepare, on short notice, for offensive capabilities rising along an exponential curve.
• Dai Zovi offered two key judgments: threat models must assume adversaries with capabilities previously reserved for nation-state actors; and enterprise defense must move beyond a purely software mindset.
• Responding to Halvar Flake's discussion of AI changing offense and defense, Green predicted a split between "fast defenders" and "slow defenders": fast defenders rapidly patch vulnerabilities and integrate AI into CI pipelines, making their software very solid, while the long tail of smaller organizations lacking such capability will keep falling behind.

Why it matters

• Green provides a simple, actionable framework: extrapolate AI capability linearly from the current pace to baseline next year's risks.
• The two experts converge on one conclusion: AI is democratizing offensive capabilities once limited to nation-states, while real-world enterprise and network defenses are far from keeping up with that curve.
• The fast/slow defender split implies AI's security impact will be uneven—top organizations may harden further with AI while the long tail's risk exposure relatively grows—giving security resourcing and policy debates a more concrete picture.

• Matthew Green: assume models keep improving at the same rate as since early this year — matthew_d_green · 2026-10-12
• Cryptographer Matthew Green: assume AI models keep improving at this year's pace — matthew_d_green · 2026-10-12
• Cryptographer Matthew Green: expect AI surprises, models enter new territory by next October — matthew_d_green · 2026-10-12
• Cryptographer Matthew Green: AI breaking crypto is the least of our security worries — matthew_d_green · 2026-10-12
• Dino Dai Zovi: AI offensive powers demand non-software defenses, OAuth tokens won't cut it — dinodaizovi · 2026-10-12
• Security vets: AI gives attackers nation-state power, and most networks aren't ready — matthew_d_green · 2026-10-12
• Cryptographer Matthew Green: AI will split software into fast and slow defenders — matthew_d_green · 2026-10-12

Episode 2 · Cryptographer Matthew Green Leads Debate: AI Bug Boom Is Patches, Not Attacks—Yet (2026-10-22, 25 posts)

On October 22, Johns Hopkins cryptographer Matthew Green posted a flurry of tweets on the topic of "AI accelerating vulnerability discovery," with a core verdict: there is no so-called "AI-driven hacking explosion" underway right now—what's actually happening is an explosion of patching, plus one old-fashioned, low-level blunder unrelated to AI. Sociologist Zeynep Tufekci shares the same view. Green quipped that if this is the new speed of vulnerability discovery, he's actually reassured—everything gets hacked eventually anyway.

Confirmed

• Green's view: the AI danger is real and an attack wave may come in the future, but right now it's a patch explosion, not an attack explosion; Zeynep Tufekci expressed the same judgment.
• Green called for building datasets: for every major vulnerability discovered in the coming months, record whether maintainers ran AI auditing tools such as Mythos or Daybreak, to measure the share of hardening fixes attributable to them and quantify the real effectiveness of AI security tools.
• Green raised another key question: does AI-generated code that heavily depends on specific models have a higher or lower vulnerability rate than human-written code? He noted that no one systematically collects such data, though AI labs should be doing exactly that.
• Green amplified an observation from security researcher joaxcar: once someone finds a vulnerability with AI, reproduction becomes trivially easy—joaxcar spent 48 hours on heap alignment to find the first GitLab RCE with Codex, while others with full information reproduced it in about 1 hour, showing the attack barrier collapsing.
• Green debated security researchers including HackingLZ over whether AI significantly lowers the barrier to lethal cyberattacks: the other side argued that groups truly intent on killing via cyber means are limited and fear retaliation, and nation-state attackers already have other avenues.

Why it matters

• The discussion pulls the focus back from panic-driven narratives to measurable empirical questions: whether AI security tools actually reduce vulnerabilities, and whether AI-generated code is less secure—both currently lack systematic data.
• The asymmetry that vulnerabilities are "hard to find first, easy to reproduce" means the attack surface expands rapidly once a flaw is exposed, making patch speed the critical line of defense.

• Nate Soares: Glasswing missing a vuln for 6+ months indicts defense-over-offense thesis — So8res · 2026-10-22
• Matthew Green debates AI lowering the bar for lethal cyberattacks, citing jailbroken GLM 6 — matthew_d_green · 2026-10-22
• Zeynep Tufekci pushes back: we see a patching explosion, not an AI hacking explosion — zeynep · 2026-10-22
• Defenders' AI, not attackers, found the vulns — fueling AI offense-defense debate — cephaloform · 2026-10-22
• Researcher: AI cut GitLab RCE reproduction from 48 hours to 1, collapsing exploit lead time — matthew_d_green · 2026-10-22
• Cryptographer Matthew Green jokes AI's bug-finding speed means 'it'll all be over in six months' — matthew_d_green · 2026-10-22
• Matthew Green asks: are bugs found with AI auditors like Mythos and Daybreak being tracked? — matthew_d_green · 2026-10-22
• Cryptographer: AI labs should track whether AI security tools actually catch real vulnerabilities — matthew_d_green · 2026-10-22
• Cryptographer asks: is AI-generated code more or less likely to have vulnerabilities? — matthew_d_green · 2026-10-22
• Security researcher: no AI hacking explosion yet, just a patching explosion — matthew_d_green · 2026-10-22
• AI cyber offense vs defense is just the same old arms race, says security pro — ziv_ravid · 2026-10-22
• Open models will be the central pillar of AI cyber defense, argues security thread on PHP bugs — ziv_ravid · 2026-10-22
• Security founder: open models will anchor AI defense, and OpenAI/Anthropic will fail to monopolize it — ziv_ravid · 2026-10-22
• Security researcher accuses AI lab figure of mischaracterizing AI-found exploit — BlancheMinerva · 2026-10-22
• Debate over open-sourcing frontier model Mythos: the php2shell exploit and offense-dominant world — Varun_Jha510 · 2026-10-22
• Blanche Minerva Accuses Nate of Cherry-Picking on AI Defense Audits — BlancheMinerva · 2026-10-22
• So8res: glasswing missing a bug for 6+ months indicts defense-beats-offense theory — harris_edouard · 2026-10-22
• Harris vs So8res: can AI-driven cyber defense ever fully match offense? — harris_edouard · 2026-10-22
• Vulnerability bonanza, yet no crypto black swan yet — and that's impressive — Kyrannio · 2026-10-23
• Security researcher: php2shell is no discontinuity, but now is the window to harden everything — joshua_saxe · 2026-10-23

5 more related posts →

Episode 3 · $10 Uncensored GLM-5.3 Sparks 'Vibe Hacking' Security Debate (2026-10-23, 8 posts)

Investor deedydas warned on Twitter on October 23 that the era of "vibe hacking" has arrived: a recent wave of critical vulnerabilities may be partly driven by the spread of "abliterated" (safety-guardrail-removed) open-source models. He noted that the platform abliterationai already offers a guardrail-free GLM-5.3—a near-frontier model—for about $10, effectively a "world-class penetration tester."

Confirmed

• An uncensored GLM-5.3 does exist on abliterationai, priced around $10 (deedydas)
• Cryptographer Matthew Green commented that uncensored GLM is merely the weakest model on the open-weight "cyberattack frontier" for the coming years; capabilities will only grow, and security risks will scale with model capability
• A shared viewpoint (from @JordanNanos): anyone sufficiently motivated, with a guardrail-free GLM-5.3 or DeepSeek V4.1 Flash 1013 model plus enough compute, can already do whatever they want
• Dan Jeffries pushed back hard: the so-called info war is stigmatizing open source; vulnerabilities have always existed, they've just been "democratized," and defensive audits uncovering them shouldn't be blamed on uncensored models

Unconfirmed

• The claim that the recent surge in critical vulnerabilities is partly due to abliterated models remains deedydas's speculative attribution with no quantitative evidence; the security community is clearly divided on it

Why it matters

• The debate touches on the core security tension of open-weight models: capabilities are rapidly approaching the frontier while removing guardrails costs as little as $10, meaning the cost of acquiring attack capability has dropped sharply
• If Green's judgment holds—uncensored models will only get stronger—defenders and policymakers will face a continually escalating open-source cyberattack frontier

• Abliterated GLM-5.3 or DS V4.1 Flash plus compute lets motivated individuals do anything, argues viral thread — khademinori · 2026-10-23
• Abliterated GLM-5.3 sells for $10 as "world-class penetration tester", fueling vulnerability surge — deedydas · 2026-10-23
• $10 Buys You an Abliterated Frontier-Class Model, and 'Vibe Hacking' Is Here — deedydas · 2026-10-23
• Deedy blames 'abliterated' open models for vulnerability wave, $10 GLM 5.3 as pen-tester — max_paperclips · 2026-10-23
• Open-model advocates push back on blames that abliterated LLMs drive vulnerability surge — max_paperclips · 2026-10-23
• Debate: are abliterated models causing more critical vulns, or just democratizing their discovery? — max_paperclips · 2026-10-23
• Matt Green: abliterated GLM is the least capable open-weight model we'll see at the cyber frontier — matthew_d_green · 2026-10-23
• ThePrimeagen: closed model gardens won't stop attackers already using abliterated models — max_paperclips · 2026-10-23

Episode 4 · Open-Source Advocates Push Back on 'Info War' Narratives Over PHP Bug (2026-10-23, 2 posts)

Dan Jeffries rebutted the 'info war' narrative against open models, clarifying that the PHP flaw was found by Sol 5.6 in defensive auditing, not by a dangerous abliterated model, urging adoption rather than fear.

• Open-model security debate: PHP bugs found by defensive audit, not abliterated models — zacharynado · 2026-10-23
• Open-model advocates push back: 'fix the buildings, don't smash the glasses' — neil_chilson · 2026-10-23

Episode 5 · Security Community Debates Whether OpenAI Waited Too Long on Dangerous Models (2026-10-24, 3 posts)

Security researchers are split over OpenAI and Anthropic's handling of dangerous cyber capabilities. Cryptographer Matthew Green argues the vendors wasted a four-month defender-only window, replying to defenders' praise with "but it was not enough."

• Matthew Green: AI labs wasted a 4-month defender-only window on cyber capabilities — matthew_d_green · 2026-10-24
• Security researchers clash over OpenAI's glasswing project and delayed cyber defense rollout — waral555 · 2026-10-24
• Matthew Green: 'And yet it wasn't enough' — matthew_d_green · 2026-10-24

Episode 6 · Cryptographer Matthew Green: Only ~5% of Vulnerable Software Fixed in 4-Month Window (2026-10-24, 2 posts)

Cryptographer Matthew Green criticized AI labs over a major security incident, noting that during a roughly four-month defensive window with near-exclusive access, only about 5% of vulnerable software was addressed.

• Matthew Green: only ~5% of vulnerable software covered during the 4-month defender window — matthew_d_green · 2026-10-24
• Matthew Green: Only ~5% of Vulnerable Software Fixed During Key 4-Month Window — waral555 · 2026-10-24