Yes—AI systems have been reported to help uncover previously unknown software vulnerabilities, including zero-days. But a model’s alert is only a lead until people verify that the flaw is real, assess its impact, and handle it responsibly. Company reports and a public-sector competition provide examples; they do not establish a universal success rate or show that AI can find every flaw.
What does “zero-day vulnerability” mean?
A zero-day is generally a vulnerability not previously known to the software maintainer or the public. The term describes the flaw’s discovery status, not its severity or what has happened to it.
As an Amazon Associate I earn from qualifying purchases.
- Finding a previously unknown flaw does not by itself prove that it can be exploited.
- It does not establish how serious the vulnerability is or whether attackers already know about or are using it.
- A suspected issue becomes useful security evidence only after it is reproduced, reviewed, and reported or fixed through an appropriate process.
The examples below are reports of previously unknown vulnerabilities or competition findings. They should not be read as evidence that each flaw was being exploited in the wild.
What evidence shows that AI can help find them?
There are concrete reported examples, but they come from specific company evaluations and a competition—not from a single independent test across the software industry.
#1 Best Overall
| Example | What was reported | What the result does and does not establish |
|---|---|---|
| OpenAI disclosure policy announcement, June 2025 | OpenAI said systems it developed had uncovered zero-day vulnerabilities in third-party and open-source software, including through automated analysis using AI tools. | This is an organizational report of findings and disclosure practice, not a measured industry-wide detection rate. |
| OpenAI Aardvark, October 2025 | OpenAI reported that Aardvark identified 92% of known and synthetically introduced vulnerabilities in “golden” benchmark repositories. It also said ten open-source findings had received CVE identifiers. | The percentage is a company-reported result on the named benchmark, not a promise about real-world detection. CVE identifiers indicate recorded vulnerabilities; the count is not a detection rate. |
| DARPA AI Cyber Challenge (AIxCC) semifinal, 2025 | DARPA reported that challenge systems found 22 unique synthetic vulnerabilities and patched 15, and found one real-world bug in SQLite3 that was responsibly disclosed. | These results came from competition systems and challenge conditions. They do not show that an AI can autonomously secure arbitrary production software. |
| OpenAI Astra update, 2026 | OpenAI reported two zero-day vulnerabilities discovered and used as part of an exploit chain in an internal evaluation, with disclosure to maintainers in progress at publication. It also described expert-led assessments that found unknown vulnerabilities in a hardened browser and operating system and formed exploit chains. | OpenAI said the Astra results reflected Daybreak Blue access rather than its default production configuration; advanced access was initially limited to a group of testers. |
| OpenAI Daybreak, August 2026 | OpenAI said GPT-5.6-Cyber was trained for specialized cybersecurity tasks, including finding zero-days and developing exploit chains. It reported using the model to investigate V8 and uncover two previously unknown vulnerabilities that researchers validated and reported to Google through coordinated disclosure. | This is a company-reported result under the stated access and evaluation conditions, not an independent comparison with other systems. |
Taken together, these examples support a measured conclusion: AI can contribute to finding previously unknown flaws, but outcomes depend on the system, task, codebase, tools, and access conditions.
How does an AI vulnerability-finding system work?
OpenAI’s description of Aardvark illustrates a repository-focused workflow. It builds a threat model from a project, reviews code changes in repository context, attempts to trigger suspected vulnerabilities in an isolated sandbox, and can propose a patch for human review.
- Understand the project: Build context about the repository and its likely security risks.
- Inspect code and changes: Look for behavior that could introduce or expose a vulnerability.
- Try to reproduce the issue safely: Attempt to trigger it in an isolated environment rather than treating a static alert as proof.
- Present evidence for review: Give a security reviewer enough information to assess the suspected behavior.
- Propose and evaluate a fix: A suggested patch still needs review and testing to make sure it addresses the flaw without breaking intended behavior.
This describes Aardvark’s reported workflow, not a universal design shared by every AI security tool. The essential distinction is between suggesting a possible flaw and producing reproducible evidence that a reviewer can verify.
Can AI detect zero-days before hackers do?
The cited examples do not establish that. A system may uncover a flaw that was previously unknown to its maintainer or the public, but that does not reveal whether a criminal or other researcher discovered it first, or whether anyone is exploiting it. The reports do not provide a general method for establishing who found a flaw first across all software.
AI-assisted discovery can give defenders another way to examine code, but no reported benchmark here proves that AI consistently beats attackers to zero-days. A finding should be treated as a vulnerability lead until it is validated and handled.
What do the reported percentages and counts mean?
Numbers from different tests are not directly comparable. They measure different things in different settings, and none supplies a cross-vendor, independently replicated success rate for finding real zero-days across software.
Rank #3
- Aardvark’s 92%: OpenAI’s October 2025 result for identifying known and synthetically introduced vulnerabilities in “golden” benchmark repositories. It is not a 92% real-world zero-day detection rate.
- Ten CVE identifiers: OpenAI said ten open-source Aardvark findings had received CVE identifiers. This is a count of reported findings, not evidence of detection performance across all projects.
- More than 40,000 CVEs in 2024: OpenAI cited this figure in its Aardvark announcement. It is OpenAI’s statement, not an independently established measure of how many flaws AI could find.
- Around 1.2% of commits introduce bugs: OpenAI reported this as a result of its testing context in the Aardvark announcement. It should not be generalized to every repository or development workflow.
- AIxCC results: DARPA reported 22 unique synthetic vulnerabilities found and 15 patched, plus one real-world SQLite3 bug found in the semifinal. These are competition outcomes, not a production-wide success rate.
Without shared datasets, conditions, and independent evaluation, a leaderboard assembled from these figures would be misleading.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWindows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCan AI write a patch for a zero-day?
It can propose a patch in some workflows, but a proposed fix is not the same as a safe, complete repair. A maintainer or security engineer needs to confirm that it closes the vulnerability, preserves intended functionality, and does not introduce a new problem.
DARPA’s AI Cyber Challenge made remediation a substantial part of the task: its final scoring algorithm gave patching while preserving functionality three times the weight of identifying vulnerabilities alone. That is a useful reminder that finding a flaw is only one step in improving security.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How should defenders validate an AI finding?
For authorized defensive work, use a repeatable process that separates the model’s suggestion from verified evidence.
- Confirm authorization and scope. Test only software and systems you own or have explicit permission to assess. A model’s output does not authorize testing someone else’s service.
- Reproduce the issue in isolation. Use a controlled test environment, such as a sandbox, and record the inputs, conditions, and behavior that trigger the suspected flaw.
- Ask a qualified reviewer to assess it. Check whether the behavior is genuinely a security problem, what an attacker could affect, and what prerequisites apply.
- Develop and test a fix. Verify that the patch addresses the reproduced issue and preserves expected behavior.
- Report it through the maintainer’s process. Share technical evidence privately where appropriate, and coordinate on remediation and disclosure.
If the issue cannot be reproduced, the report remains unverified; do not present it as a confirmed zero-day.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Who can use these capabilities?
Availability depends on the product, access tier, and safeguards. OpenAI’s Daybreak announcement describes Blue access for approved defensive work and Red access for authorized vulnerability research, exploit validation, and security testing. OpenAI also says GPT-5.6-Cyber is intended for specialized cybersecurity tasks.
Best Value
OpenAI said Astra’s reported results reflected Daybreak Blue access, not its default production configuration, and that enhanced checks could slow, pause, or stop legitimate work. A capability demonstrated in an internal evaluation or access-limited setting should not be assumed to be available in a public chatbot or to produce the same result for every user.
What does responsible disclosure involve?
When a finding appears real, it should be communicated to the maintainer or vendor so the issue can be assessed and addressed. In its June 2025 announcement, OpenAI described a process involving validation, prioritization, vendor contact, and non-public disclosure by default. OpenAI said its default timeline was open-ended rather than a fixed universal deadline, while reserving the option to disclose in some circumstances, such as public interest.
That is OpenAI’s stated approach, not a rule that applies to every company or vulnerability. Follow the affected maintainer’s published reporting process, and avoid publicly releasing exploit details before coordinating with the people responsible for the software.
Recommended Free Tools
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
