How do OpenAI, Microsoft, Anthropic and Google AI bug hunters compare in November 2026?
OpenAI, Microsoft, Anthropic and Google DeepMind all reported AI-found vulnerabilities in three weeks. What each counted, what it cost and what it means.
By Emre A.Published
Tech AI LatestIn three weeks, four labs published AI vulnerability research: OpenAI validated 948 Linux kernel flaws for about $1.2 million, Anthropic chained bugs into Chrome and Linux exploits, Microsoft turned 26,502 findings into 59 kernel reports and 33 CVEs, and Google DeepMind weakened 1024-bit RSA. The numbers measure different things. Microsoft's data shows the real limit: finding bugs takes weeks, but validating and fixing them takes people months.
Key facts
- OpenAI: 948 validated Linux kernel privilege escalation flaws in one run, about $1.2 million in API credits (4 November)
- Anthropic: exploit chains in Chrome, Safari, Android, Windows and Linux; 62 Chrome CVEs (27 October); Xen XSA-523 credited to Claude (3 November)
- Microsoft: 26,502 raw findings, 621 proven cases, 59 reports, 33 CVEs, 8 root exploits, 38 merged fixes (11 November)
- Google DeepMind: two cryptanalysis results against legacy 1024-bit RSA; 2048-bit RSA not affected, it says (6 November)
- Only OpenAI disclosed a cost; only Microsoft disclosed how long fixes took to merge
Why this question is in the news
Between 27 October and 11 November 2026, four of the largest AI developers published reports on vulnerabilities their models found in real software. Each came with a large number, and the numbers are easy to line up side by side. They should not be, at least not without reading what each one counts. This explainer sets out what each lab reported, what the figures measure, what they cost where that is known, and what the reports mean for companies that run the affected software.
The background is a measurable rise in disclosures. Google Threat Intelligence Group reported on 30 September that monthly vulnerability disclosures doubled from 5,045 in January 2026 to 10,740 in August, and that 50% of bugs found by AI lead to remote code execution, against 26% across all CVEs. The lab reports of the past three weeks are the supply side of that curve: they describe how the bugs are being found.
What each lab reported
OpenAI. On 4 November, OpenAI described a pipeline in which agents on an unreleased cyber model analysed 2,308 candidate Linux kernel findings, reproduced 1,724 and validated 948 local privilege escalation vulnerabilities, each with a candidate patch, in a single run. OpenAI put the cost at about $1.2 million in API credits, roughly $1,300 per validated flaw. A day later it released GPT-6 Cyber to approved Daybreak Red customers and said the model found 81 previously unknown high-severity Firefox flaws. OpenAI reports a score of 48.5% on its ExploitGym benchmark, against 32.1% for GPT-5.6 Cyber, and rates the model Critical for cybersecurity under its Preparedness Framework.
Anthropic. On 27 October, Anthropic's Frontier Red Team published five exploit chains built by an unreleased internal model: one-click remote code execution in Chrome, a Safari same-origin bypass, a Pixel 9 lock screen bypass, a Windows kernel remote denial of service and a zero-click Linux root exploit in multipath TCP, CVE-2026-98432. Google assigned 62 CVEs, all critical or high, fixed in Chrome 156. A week later the Xen Project's advisory XSA-523 credited Anthropic, using Claude, for two flaws that let a malicious guest read and write memory of other virtual machines. Anthropic gave no cost or total time.
Microsoft. On 11 November, Microsoft's FORGE Lab reported on its MDASH system. Scans of selected Linux kernel subsystems produced 26,502 raw findings in about two weeks; 621 cases had proof of concept; 59 became public reports. Those led to 33 CVEs, 24 rated High or above, 8 confirmed root exploits and 38 fixes merged into mainline. Microsoft used models from more than one vendor: Sonnet 4.6, GPT-5.4 and GPT-5.5 for scanning, and GPT-5.5, Opus 4.6 and GPT-5.5-cyber to reproduce bugs.
Google DeepMind. On 6 November, DeepMind said agents guided by its researchers produced two cryptanalysis results against legacy 1024-bit RSA: a signature forgery attack it estimates at about 1,600 times better than the best known method in one setting, and a batch factoring method that cuts the cost per extra key about 30 times. DeepMind says 2048-bit RSA is not affected. This is a different kind of result: a mathematical weakening of an old standard, not a list of bugs in code.
Smaller firms report the same pattern at smaller scale. CISA credited Causal Security's autonomous system Cipher with four memory-safety flaws in EPICS Base, the control software used at particle accelerators and telescopes, one rated 9.8; Causal Security says Cipher found all four in 3 hours and 49 minutes.
Why the numbers do not compare directly
The headline figures count different stages of the same process. OpenAI's 948 is a count of validated vulnerabilities with candidate patches, before upstream review; the company did not list CVE numbers or affected versions. Microsoft's 59 is a count of reports it chose to file after removing duplicates and low-impact cases, and its 33 is the number of CVEs assigned so far. Anthropic's 62 Chrome CVEs were assigned by Google, the vendor, which makes them the most externally confirmed figure of the set, but they come from one attack campaign on one product.
Microsoft's funnel shows how much the stage matters. Of its 621 cases with working proof of concept, 266 shared a root cause with another case, 163 had limited security impact and 133 were already reported or fixed. Only 9.5% became reports. Applied to any raw count, a ratio like that would shrink the headline sharply, although there is no reason to assume other pipelines filter at the same rate. Coverage of OpenAI's post also differs on whether 948 counts distinct vulnerabilities after merging duplicates or working exploits, and OpenAI has not published a breakdown like Microsoft's.
Cost is the other gap. OpenAI is the only lab that published a price, about $1.2 million for one run. Microsoft gave compute in servers and weeks, two servers for about two weeks of scanning, plus people: one full-time researcher and one intern over three months. Anthropic and DeepMind gave neither. For readers comparing model prices, the AI model API pricing guide explains how per-token prices turn into project costs, but no lab has yet published enough to compute a cost per fixed bug.
The real bottleneck: validation and fixes
The most practical finding in the set is Microsoft's account of time. Reporting 59 bugs took more than four person-months. The 38 merged fixes took about five months to land; half were accepted on first submission and the rest needed one to four revisions. The median time from first public patch to a mainline kernel release was 41 days, and the longest was 108. Microsoft's conclusion is that more tokens find more bugs but do not by themselves make software secure.
Kernel maintainers have said the same from the other side. The Linux kernel's own security bug documentation warns that AI-assisted bug hunting has flooded maintainers with poor-quality reports and asks reporters to keep submissions short and accurate. OpenAI says it sends the most critical bugs first and routes the rest through its Patch the Planet partners; Anthropic uses a 90-day disclosure window plus 45 days before discussing details. Both are attempts to stop discovery from overwhelming the people who merge fixes.
There is also an access question. The strongest cyber models are restricted. GPT-6 Cyber is limited to approved Daybreak Red customers, and Anthropic offers reduced blocking for qualifying security professionals through its Cyber Verification Program. On 12 November OpenAI added Daybreak Red Team and Defense Factory, products that test web applications and prepare fixes for engineers to review, without published prices or accuracy figures. Defenders outside these programs work with general models, and open-weight models carry no such access gate at all; the open-weight models guide explains how those are released and licensed.
What this means for companies
For most organisations the reports change patch planning more than tooling. The Linux bugs in the OpenAI and Microsoft reports are mainly local privilege escalation flaws: they do not let an attacker in from the internet, but they turn any foothold, such as a compromised web application, a malicious container or a CI job, into root on the host. Shared servers, build runners and multi-tenant clusters carry the most risk, and their kernels should be updated on the distribution's security schedule rather than at the next maintenance window.
Three checks follow from the numbers. First, expect a higher rate of kernel and browser CVEs into 2027 as these batches move through disclosure, and plan patch capacity for it. Second, treat vendor counts as claims until CVEs and advisories appear; the gap between Microsoft's 621 cases and 59 reports shows how much a raw count can shrink. Third, if a team buys an AI security tool, ask for the same metrics Microsoft published: how many findings were real, how many were duplicates, and how long fixes took. More background is in the site's guides and the AI cybersecurity section.
Questions people ask
- Which AI lab has found the most vulnerabilities?
- OpenAI reports the largest single count, 948 validated Linux kernel flaws in one run, but the figures from OpenAI, Microsoft and Anthropic count different stages, so they cannot be ranked directly.
- How much does AI vulnerability hunting cost?
- Only OpenAI published a figure: about $1.2 million in API credits for its 948-flaw kernel run, roughly $1,300 per validated flaw. Microsoft, Anthropic and Google DeepMind gave no cost.
- Are the AI-found Linux kernel bugs dangerous for servers?
- Most are local privilege escalation flaws. An attacker needs to run code on the machine first, so shared hosts, CI runners and container platforms are most exposed. Updated distribution kernels fix them.
- Can anyone use the strongest AI cyber models?
- No. GPT-6 Cyber is limited to approved OpenAI Daybreak Red customers, and Anthropic's stronger cyber settings require its Cyber Verification Program.
Sources
- Lessons from 59 Linux Kernel Bug Reports (Microsoft, 11 Nov 2026)
- A thousand fewer paths to root (OpenAI, 4 Nov 2026)
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (OpenAI, 5 Nov 2026)
- Discovering and exploiting vulnerabilities in critical, real-world software with Claude (Anthropic, 27 Oct 2026)
- XSA-523: Arbitrary read/write via grant copy (Xen Project, 3 Nov 2026)
- Advancing the frontier of cryptanalysis (Google DeepMind (Pushmeet Kohli), 6 Nov 2026)
- Vulnerability Discovery and Exploitation Trends in the AI Era (Google Threat Intelligence Group, 30 Sep 2026)
- ICSA-26-300-01: EPICS Base (CISA, 27 Oct 2026)
Reference pages
- GPT-6 Cyber model
- OpenAI company
- Microsoft company
- Anthropic company
- Google DeepMind company
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.





