Venture fundsFund closes, LPs & managers
a16z's Zane Lackey Says AI-Found Zero-Days Are Bringing Back Early-Internet Security Risks
An a16z essay by Zane Lackey argues AI models now find serious software flaws faster than maintainers can fix them, echoing the zero-day era of the late 1990s and early 2000s, and predicts returning worms, faster patching and rebuilt security stacks.
From the a16z article of November 16, 2026 and cited sources: AI is finding serious bugs faster than they can be fixed. [2] [5]
SOURCES[2] Anthropic: Glasswing initial update anthropic.com[5] Anthropic: LLM-discovered 0-days anthropic.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- Original source
- a16z article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 16, 2026 article published
- Covered
- Nov 16 edition
- Evidence
- 17 sources across 14 sites
- 1The Dream of the 1990s Is Alive Again in Cybersecuritya16z.news
- 2Anthropic: Glasswing initial updateanthropic.com
- 3Senate HSGAC hearing summary, May 19, 1998webarchive.library.unt.edu
- 4Notre Dame: Hackers' testimonyethicsatwork.nd.edu
- 5Anthropic: LLM-discovered 0-daysanthropic.com
- 6Axios: Opus 4.6 security flawsaxios.com
RisingHow it’s measured
Substack · 158 likes · 1 comment
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

AI is finding serious bugs faster than they can be fixed. Partners in Anthropic's Project Glasswing found more than 10,000 high- or critical-severity bugs, but only 75 of the 530 reported had been patched. Anthropic reports these totals itself. In the one independent check, outside security firms reviewed 1,752 findings and judged 90.6% valid and 62.4% high or critical. [2] [5]
Lackey compares today to 1998, when L0pht hackers told a Senate panel they could take down the internet in 30 minutes. He argues attackers now work at machine speed. He predicts internet worms will return, patch cycles will shrink sharply and security will move from yearly compliance pentests to continuous red-team testing with no scope limits. [1] [3] [4]
Attacks already tend to come before fixes. The Cloud Security Alliance cites an average time-to-exploit of minus seven days. It also reports that nearly 29% of the flaws on CISA's 2025 known-exploited list were exploited on or before their CVE was published. Its advice is runtime detection, zero trust and ranking flaws by exploit probability, not severity score alone. [11]
Glasswing partners found over 10,000 serious bugs, yet few of those reported are patched, and exploitation frequently begins before disclosure.
Read the full assessment
Implication: defenders may need runtime detection, exploit-probability prioritization and continuous testing rather than periodic patch cycles.
Executive brief
AI models are now finding serious bugs faster than maintainers can fix them. Anthropic's own Project Glasswing update says that of 530 high- or critical-severity bugs it reported, only 75 had been patched. In an a16z newsletter post, Zane Lackey of a16z compares this to the early 2000s, when a steady stream of zero-days hit core internet software. He expects internet worms to return, patching to become much faster and security stacks to be rebuilt. He also recommends several companies a16z has invested in. The post is commentary. The data behind it comes mostly from AI vendors.
What changed and event timeline
Opus 4.6 finds 500+ zero-days
Anthropic reported more than 500 validated high-severity open-source bugs, including in Ghostscript, OpenSC and CGIF. It said the model found them with standard tools and no special setup (;).
Project Glasswing launches
Claude Mythos Preview was made available only to selected partners, including AWS, Apple, Google, Microsoft and the Linux Foundation. Anthropic committed $100M in usage credits ().
Patching becomes the bottleneck
Partners had found more than 10,000 high- or critical-severity bugs. Only 75 of 530 reported had been patched ().
Neo raises $100M
Neo launched to secure AI agents running on company devices. a16z and Bessemer led the round (;).
a16z publishes its thesis
Lackey argues that offense now runs at machine speed. He predicts worms, faster patching and hardened infrastructure products ().
Capabilities and access
- Claude Opus 4.6: generally available. Anthropic added cyber-specific monitoring and can block traffic it flags as malicious (Anthropic).
- Claude Mythos Preview: not publicly released. Approved partners reach it through the Claude API, Bedrock, Vertex AI and Microsoft Foundry.
- Hinoki Security: the post says it offers security-tuned models that don't refuse red-team tasks.
Read the full section
- Claude Opus 4.6: generally available. Anthropic added cyber-specific monitoring and can block traffic it flags as malicious (Anthropic).
- Claude Mythos Preview: not publicly released. Approved partners reach it through the Claude API, Bedrock, Vertex AI and Microsoft Foundry. After the credits run out it costs $25 per million input tokens and $125 per million output tokens (Anthropic Glasswing).
- Hinoki Security: the post says it offers security-tuned models that don't refuse red-team tasks. Its public profile describes a platform for prioritizing and fixing vulnerabilities (Built In).
Technical analysis for researchers and developers
- How Anthropic ran it: Opus 4.6 worked in a sandboxed VM with debuggers and fuzzers.
- Independent checks on Glasswing: outside security firms reviewed 1,752 findings. 90.6% were valid, and 62.4% were confirmed high or critical (Glasswing update).
- Reproducibility: the full lists of vulnerabilities are held back under coordinated disclosure, so outsiders cannot yet replicate the results.
Read the full section
- How Anthropic ran it: Opus 4.6 worked in a sandboxed VM with debuggers and fuzzers. Anthropic focused on memory-corruption bugs because crashes and address sanitizers make them easy to confirm. Human researchers checked each finding and wrote the patches (Anthropic).
- Independent checks on Glasswing: outside security firms reviewed 1,752 findings. 90.6% were valid, and 62.4% were confirmed high or critical (Glasswing update).
- Reproducibility: the full lists of vulnerabilities are held back under coordinated disclosure, so outsiders cannot yet replicate the results.
Claims and evidence
- "500+ zero-days" and "10,000+ high/critical": Anthropic's own numbers.
- Exploitation starts before patches exist: the Cloud Security Alliance cites M-Trends 2026 for an average time-to-exploit of −7 days.
- Worms returning and attackers stealing inference API keys: Lackey's forecast. No independent reporting confirming either was found.
Read the full section
- "500+ zero-days" and "10,000+ high/critical": Anthropic's own numbers. The 90.6% validity rate is the only independent check, and it covers a subset of findings (Glasswing update).
- Exploitation starts before patches exist: the Cloud Security Alliance cites M-Trends 2026 for an average time-to-exploit of −7 days. It also reports that 28.96% of the vulnerabilities on CISA's 2025 known-exploited list were exploited on or before their CVE was published (CSA).
- Worms returning and attackers stealing inference API keys: Lackey's forecast. No independent reporting confirming either was found.
Context and prior work
- Early AI exploitation research showed GPT-4 exploiting 87% of tested one-day vulnerabilities (already disclosed, not yet widely patched), at about $8.80 per exploit (CSA).
- Anthropic separately reported 22 Firefox bugs found by Opus 4.6. Mozilla later fixed 271 bugs in Firefox 150 using Mythos (Anthropic Firefox; Glasswing update).
Limitations, safety and contested findings
- Bruce Schneier amplified the Opus 4.6 claims.
- Anthropic says the industry-standard 90-day disclosure window may not hold up at this speed and scale (Anthropic).
- Models tuned not to refuse offensive tasks need very tight deployment controls (DEV Community).
Read the full section
- Bruce Schneier amplified the Opus 4.6 claims. Commenters on his post noted that only three bugs were described in detail, and argued the model mostly finds new examples of known bug types rather than new kinds of attack (Schneier).
- Anthropic says the industry-standard 90-day disclosure window may not hold up at this speed and scale (Anthropic).
- Models tuned not to refuse offensive tasks need very tight deployment controls (DEV Community).
Business and practitioner implications
- Patching: the CSA puts average enterprise remediation for complex applications at about five months and ten days.
- Testing: Lackey urges continuous red-team testing with no scope limits, rather than an annual pentest for compliance.
- Disclosure of interest: the post recommends Neo, Keycard and Socket, all a16z investments.
Read the full section
- Patching: the CSA puts average enterprise remediation for complex applications at about five months and ten days. It recommends runtime detection, zero trust and prioritizing by EPSS exploit probability instead of severity scores alone (CSA).
- Testing: Lackey urges continuous red-team testing with no scope limits, rather than an annual pentest for compliance.
- Disclosure of interest: the post recommends Neo, Keycard and Socket, all a16z investments. Their funding:
- Neo: $100M across seed and Series A, led by a16z and Bessemer. Company-announced and independently reported; valuation not disclosed (Fortune).
- Keycard: $38M in total. SecurityWeek says Acrew led the $30M Series A and a16z co-led the $8M seed. Other outlets credit a16z with leading, so the accounts conflict (SecurityWeek).
- Socket: $60M Series C at a $1B valuation, as announced by the company (Socket).
- Hinoki: a16z backing appears only in a recruiting profile (Vaia).
The source trail.
Sources (17)
The Dream of the 1990s Is Alive Again in Cybersecurity
Article text retrieved; extracted text may omit tables or interactive elements.
a16z.news