Can AI Really Find Zero-Day Vulnerabilities Before Humans?
For decades, discovering serious software vulnerabilities has depended heavily on human security researchers. Researchers read source code, understand complicated program behavior, build testing environments, analyze crashes, investigate unusual execution paths, and spend days or sometimes months trying to determine whether an obscure bug can become a meaningful security problem.
Artificial intelligence is beginning to change that process.
Modern frontier models can analyze enormous amounts of code, reason about program behavior, identify suspicious patterns, generate tests, inspect previous fixes, and repeatedly investigate promising areas. The most important development is not that AI can identify ordinary coding mistakes. Automated security tools have been doing that for years. The important change is that advanced models are increasingly capable of finding previously unknown vulnerabilities, including vulnerabilities in heavily tested software.
Anthropic reported in February 2026 that Claude Opus 4.6 could find high-severity vulnerabilities in well-tested codebases, including flaws that had survived extensive automated testing and years of human review. The company warned that AI systems could eventually discover novel vulnerabilities faster and at greater scale than expert human researchers. (https://www.anthropic.com/research/zero-days)
Later in 2026, Anthropic's Project Glasswing described a more advanced frontier model, Claude Mythos Preview, being used to identify thousands of vulnerabilities across critical software. Anthropic said the model had found vulnerabilities in major operating systems and web browsers and was being made available through a controlled defensive initiative rather than general public release. (https://www.anthropic.com/glasswing)
OpenAI has also reported that its GPT-6 Astra reached the company's Critical cybersecurity capability threshold. OpenAI says that, with appropriate tools and access, Astra can find previously unknown security flaws and develop exploitation methods across many well-protected systems without a person guiding every step. (https://openai.com/index/path-to-astra/)
These developments raise a remarkable question: could an AI system discover a vulnerability before any human researcher realizes that the vulnerability exists?
The answer is increasingly yes in controlled environments and certain software-development contexts, but the reality is more nuanced than the phrase āAI hackersā suggests. AI does not automatically outperform humans at every security task. Current systems still struggle with long, complex operations, unusual architectures, incomplete context, and tasks requiring deep specialist knowledge.
What is changing is the economics and speed of vulnerability research.
What Is a Zero-Day Vulnerability?
A zero-day vulnerability is a security weakness that is unknown to the party responsible for fixing or defending the affected software, or for which no effective fix is yet broadly available. The term is commonly associated with the limited time defenders have to respond once the vulnerability becomes known or exploited.
The phrase is often misunderstood.
A zero-day is not necessarily an extremely complicated bug. A simple programming mistake can become a zero-day if nobody has previously discovered it. Conversely, a technically sophisticated vulnerability may have existed for years but not qualify as a zero-day once it has been publicly identified and addressed.
Zero-day discovery therefore depends on novelty.
A security scanner looking for known vulnerability patterns may be extremely effective but is not necessarily discovering zero-days. A researcher or AI system that identifies an entirely new security flaw is operating at a different level of the vulnerability-research problem.
Why Zero-Day Discovery Is So Difficult
Modern software is extraordinarily complex.
A browser can contain millions of lines of code. Operating systems contain enormous numbers of interacting components. Cloud applications depend on databases, APIs, authentication systems, operating-system services, libraries, containers, identity systems, and external infrastructure.
A vulnerability can emerge from the interaction between individually reasonable components.
That means simply searching for suspicious lines of code is not enough.
A researcher may need to understand how data moves through multiple functions, how permissions change, what happens after an error, how memory is managed, or how several services communicate.
Human researchers are exceptionally good at understanding context and forming hypotheses. But they have finite time and attention.
AI systems can potentially investigate many more possibilities in parallel.
How Human Researchers Traditionally Find New Vulnerabilities
Security research usually combines several techniques.
Researchers may perform manual code review, fuzzing, static analysis, dynamic analysis, reverse engineering, symbolic reasoning, differential testing, historical patch analysis, threat modeling, and controlled penetration testing.
One particularly powerful human skill is recognizing patterns.
A researcher might notice that a recent security fix addressed one variation of a programming mistake while another similar code path remains vulnerable.
Another researcher might understand that an apparently harmless input becomes dangerous only after passing through several layers of processing.
This type of reasoning is difficult to completely automate with traditional tools.
AI changes the equation because large language models can reason over code and natural-language documentation simultaneously.
Why AI Is Different From Traditional Vulnerability Scanners
A conventional vulnerability scanner generally works from predefined knowledge.
It may search for known vulnerable versions, insecure configurations, recognizable code patterns, exposed services, or established vulnerability signatures.
AI can potentially operate more like a research assistant.
It can inspect code, formulate a hypothesis, examine related functions, compare implementations, reason about possible execution paths, and decide whether additional investigation is justified.
This does not make AI infallible.
It can hallucinate vulnerabilities, misunderstand code, produce false positives, overlook important context, or incorrectly assess severity.
However, when an AI model is combined with specialized security tools, automated testing, source-code access, and carefully designed evaluation environments, it can become a powerful vulnerability-discovery system.
Anthropic's Evidence of AI-Discovered Zero-Days
One of the clearest examples came from Anthropic's February 2026 research on LLM-discovered zero-days.
Anthropic reported that Claude Opus 4.6 could identify meaningful vulnerabilities in heavily tested codebases without requiring highly specialized scaffolding for every discovery. The company said some vulnerabilities had remained undetected for years despite extensive testing. (https://www.anthropic.com/research/zero-days)
This is important because the comparison is not simply between AI and an inexperienced programmer.
The relevant comparison is between AI-assisted vulnerability research and mature security processes that may already include fuzzers, automated testing, code review, and years of human examination.
If an AI system identifies a flaw that survived those defenses, it demonstrates that AI can add a new layer of reasoning to existing security infrastructure.
Project Glasswing and the Search for Vulnerabilities at Scale
Anthropic later expanded this concept through Project Glasswing.
In April 2026, Anthropic announced a coalition involving major technology and security organizations to use its Claude Mythos Preview for defensive vulnerability research. Anthropic described Mythos as a frontier model capable of discovering vulnerabilities across critical software and said the system had already found thousands of high-severity vulnerabilities, including issues in major operating systems and browsers. (https://www.anthropic.com/glasswing)
The project is significant because it illustrates a possible future workflow:
AI discovers ā researchers validate ā maintainers are notified ā vulnerability is fixed ā AI checks the fix.
That cycle could operate much faster than conventional vulnerability research.
Instead of using AI primarily as a chatbot for security professionals, organizations can use it as a continuous research system examining software before attackers discover weaknesses.
Could AI Find a Vulnerability Before a Human?
There are several circumstances where this is increasingly plausible.
The first is scale.
An AI system can analyze large amounts of source code continuously. A human researcher cannot manually inspect every function in millions of lines of software every day.
The second is persistence.
An agent can continue investigating a promising hypothesis without becoming tired or needing to switch tasks.
The third is pattern comparison.
An AI model can compare large numbers of functions and historical changes to identify recurring structures that might deserve investigation.
The fourth is tool orchestration.
An agent can potentially combine code analysis, testing, documentation retrieval, debugging, and vulnerability databases into one workflow.
These characteristics do not guarantee that AI will discover every vulnerability first. They simply create circumstances in which AI can investigate more possibilities than an individual human can reasonably examine.
Why āAI Will Replace Security Researchersā Is Too Simple
The opposite extreme is also misleading.
Security research involves creativity, domain expertise, judgment, and contextual understanding that remain difficult for AI systems.
The UK's National Cyber Security Centre has noted that frontier AI models still have important limitations in complex cybersecurity operations. Earlier-generation systems struggled with reverse engineering, cryptography, malware development, concurrent processes, long-running operations, and consistency across repeated attempts. (https://www.ncsc.gov.uk/sites/default/files/2026-03/Why-cyber-defenders-need-to-be-ready-for-frontier-AI.pdf)
These limitations matter.
A model may identify a suspicious function but fail to understand the larger architecture. It may correctly notice an unusual memory operation but incorrectly conclude that it is exploitable. It may identify a theoretical attack path that cannot work under the actual deployment conditions.
Human validation remains essential.
The likely future is therefore not āAI replaces every security researcher.ā It is more accurately described as AI increasing the amount of vulnerability research that security researchers can perform.
The New Race Is About Speed
The cybersecurity race is increasingly becoming a race against time.
Suppose a vulnerability takes a human team three weeks to discover, reproduce, understand, and report.
If an AI system can identify the same class of weakness in hours, the discovery timeline changes dramatically.
Now consider what happens if attackers gain similar capabilities.
A vulnerability that previously remained obscure for years might be discovered shortly after it is introduced into software.
This creates pressure on developers to detect weaknesses earlier, security teams to monitor continuously, and software maintainers to patch faster.
The question is no longer simply whether a vulnerability exists.
The question becomes: how long can it remain undiscovered?
Why AI Could Search for Bugs Continuously
Human researchers generally work in projects.
A vulnerability researcher might spend several days analyzing one application, then move to another task.
An AI-driven system can potentially run continuously.
It can examine a new software release, compare it against previous versions, identify changed components, prioritize suspicious modifications, and investigate them automatically.
This creates the possibility of an always-on vulnerability research layer.
Google has already described defensive systems such as Big Sleep that use AI to detect vulnerabilities, while CodeMender uses AI reasoning to help identify and automatically fix security problems in code. (https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/google-threat-intelligence-group-report/)
The long-term implication is significant: vulnerability research could become part of the normal software development lifecycle rather than an occasional specialist activity.
AI Can Analyze Code Humans Rarely Have Time to Read
Large software projects contain enormous amounts of code that receives little direct human security attention.
Developers may understand the parts they maintain, while security teams concentrate on exposed components and known risk areas.
Deep dependencies may receive less scrutiny.
AI can potentially examine these neglected areas at scale.
This does not mean every line needs to be analyzed with the same intensity. An intelligent system can prioritize sections based on complexity, privilege, external input, historical changes, and security relevance.
That prioritization itself can become a major productivity advantage.
Why Historical Patches Are Valuable to AI
Security patches contain information about how vulnerabilities occur.
A patch can show which code changed, what assumptions were incorrect, and what behavior developers considered dangerous.
An AI system can potentially study previous fixes and search for similar patterns elsewhere in a project.
This is one reason language models are interesting for vulnerability discovery.
They can connect source code with natural-language explanations, commit histories, documentation, bug reports, and security advisories.
Traditional tools are excellent at mechanical analysis. AI can add contextual reasoning.
The Problem of False Positives
AI vulnerability discovery has an important weakness: finding something suspicious is not the same as finding a real vulnerability.
A model can identify code that looks dangerous but is protected by another layer of validation.
It can misunderstand a security assumption.
It can mistake unreachable code for exploitable code.
It can overestimate the impact of a theoretical weakness.
Therefore, AI-generated vulnerability reports require validation.
The strongest systems combine model reasoning with automated tests and independent verification rather than trusting the model's conclusion by itself.
Why Combining AI With Fuzzing Is Powerful
Fuzzing has been a major cybersecurity technique for decades.
A fuzzer generates many inputs and observes how software behaves. It is particularly effective at finding crashes and unexpected states.
AI can potentially make fuzzing more intelligent.
Instead of generating inputs blindly, an AI system can study the code and identify areas that deserve deeper testing. It can reason about likely input structures, interpret crashes, group similar failures, and determine which observations deserve further investigation.
This creates a hybrid architecture.
AI provides reasoning while specialized security tools provide evidence.
That combination is likely to be more reliable than either technology alone.
Why Zero-Day Discovery Is a Dual-Use Capability
Finding vulnerabilities is beneficial when the goal is to protect systems.
The same capability can become dangerous when used to compromise systems without authorization.
This is the central dual-use problem in AI cybersecurity.
A model that discovers a previously unknown flaw can help a software maintainer fix it before attackers learn about it.
But if the same capability is connected to unauthorized targets and offensive tooling, it could accelerate malicious activity.
This is why frontier cybersecurity models are increasingly being released with restrictions, monitoring, controlled access, and dedicated defensive programs.
OpenAI's Astra Shows How the Capability Frontier Is Moving
OpenAI's September 2026 safety documentation provides another important data point.
OpenAI says GPT-6 Astra is its first model to reach the Critical cybersecurity capability level under its Preparedness Framework. The company defines that level partly around the ability to identify and develop functional zero-day exploits against hardened real-world systems without human intervention, or to execute novel end-to-end cyberattack strategies from a high-level goal. (https://openai.com/index/path-to-astra/)
This does not mean Astra can freely compromise arbitrary systems on the public internet.
Capability evaluations are performed under specific conditions, and real-world environments introduce many additional barriers.
The significance is that leading AI developers now consider autonomous vulnerability discovery and advanced cyber operations serious enough to require dedicated safety thresholds.
Why Human Security Researchers Still Have an Advantage
Human experts possess several advantages that AI systems continue to struggle with.
Humans can interview developers, understand organizational context, reason about business consequences, make nuanced judgments about risk, and recognize when technical findings do not make practical sense.
A human researcher can also decide that a seemingly minor issue is important because it interacts with a particular business process.
AI systems may not have access to that context.
Human researchers also remain responsible for ethical boundaries and coordinated disclosure decisions.
Therefore, the strongest model is likely to be human-AI collaboration rather than complete replacement.
Could AI Discover Vulnerabilities in Software Nobody Is Watching?
This may be one of the most important consequences.
Popular software receives significant security attention. Less popular libraries may not.
Open-source projects can have relatively small maintenance teams even when their components are embedded inside thousands of applications.
An AI system capable of continuously analyzing open-source software could potentially discover vulnerabilities that would otherwise remain unnoticed.
Palo Alto Networks' Unit 42 recently described an autonomous vulnerability research system called NOVA. The company reported that over a two-month period its system analyzed 3,915 open-source projects and identified 14,090 confirmed vulnerabilities, with 99.4% described as previously unreported. The findings were being handled through responsible disclosure processes. (https://unit42.paloaltonetworks.com/frontier-ai-vulnerability-burst/)
These figures come from a vendor's own research system and should be interpreted in that context, but they illustrate the scale at which AI-assisted vulnerability research can operate.
What Happens When AI Finds More Vulnerabilities Than Humans Can Fix?
This is a less obvious problem.
Discovery is only the first stage.
Once vulnerabilities are identified, developers must reproduce them, determine severity, create patches, test the patches, release updates, and help users deploy them.
If AI increases vulnerability discovery dramatically while human remediation capacity remains unchanged, organizations could face a new bottleneck.
The industry may therefore need automated remediation systems as well as automated discovery.
AI could eventually participate in the entire defensive lifecycle:
- Identify suspicious code.
- Generate a vulnerability hypothesis.
- Validate the issue in an isolated environment.
- Classify severity.
- Propose a patch.
- Run regression tests.
- Submit the patch for human review.
- Monitor the released fix.
Human oversight remains especially important for high-impact changes.
AI Could Make Secure Software Development More Continuous
Traditional application security often occurs at specific stages.
Developers write code, automated tests run, security tools scan the project, and specialists review significant findings.
AI agents could make security analysis continuous.
Every meaningful code change could trigger an intelligent review. The agent could compare the change against historical vulnerabilities, inspect related functions, test suspicious behavior, and explain its findings to developers.
This could reduce the time between vulnerability introduction and vulnerability discovery.
In an ideal defensive system, the vulnerability is discovered by an AI security agent before the software reaches production.
The Attack Surface Is Also Growing
AI-generated software introduces another complication.
As developers increasingly use AI coding assistants, more software is being produced with AI involvement.
That means AI is simultaneously becoming a tool for creating software and a tool for examining software.
This creates a feedback loop.
AI-generated code can contain subtle mistakes. Other AI systems can potentially identify those mistakes. Attackers can also use AI to search for the same weaknesses.
The result is a faster software-security lifecycle.
Could AI Find a Zero-Day Before the Developer Who Wrote the Code?
In some situations, yes.
A developer generally focuses on whether the software performs its intended function.
A security researcher asks how the software behaves when assumptions are violated.
An AI security system can potentially combine both perspectives.
It can understand the intended function from documentation while simultaneously searching for unusual states and unexpected interactions.
This is particularly valuable in large projects where no individual developer understands every dependency.
Why Complexity Favors Automated Analysis
Software complexity increases faster than the number of security experts available to inspect it.
Modern applications depend on cloud services, APIs, open-source packages, browser frameworks, databases, identity providers, container systems, and third-party integrations.
A vulnerability may emerge at the boundary between these components.
AI agents can potentially analyze these relationships faster than humans working manually.
This makes graph-based security reasoning increasingly important.
Instead of examining vulnerabilities as isolated bugs, AI systems can reason about relationships among code, identities, services, data, and permissions.
Why AI Vulnerability Discovery Does Not Mean Automatic Exploitation
Finding a vulnerability and successfully exploiting it are different tasks.
A vulnerability may exist but require a very specific configuration. It may require authentication. It may be unreachable from an external network. It may depend on a particular version or deployment condition.
Therefore, a vulnerability report must be interpreted in context.
This distinction is important because sensational descriptions of AI cybersecurity often imply that discovering a bug automatically means compromising the system.
It does not.
The practical security impact depends on exploitability, exposure, privileges, mitigations, environmental conditions, and the value of the affected asset.
Why Defenders Need AI Too
If attackers can search software continuously, defenders cannot depend entirely on periodic manual audits.
AI can help defenders maintain a comparable operational tempo.
Security teams can use AI to inspect code, prioritize vulnerabilities, analyze alerts, monitor changes, search threat intelligence, identify suspicious behavior, and assist with incident response.
Google's threat intelligence research has already described AI being used for both offensive observations and defensive vulnerability discovery. Google reported a case in which a threat actor used an AI-assisted process to develop a zero-day exploit, while Google simultaneously uses systems such as Big Sleep to discover vulnerabilities defensively. (https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/google-threat-intelligence-group-report/)
This is the emerging competition: AI-assisted offense versus AI-assisted defense.
The Importance of Responsible Disclosure
When an AI discovers a previously unknown vulnerability, responsible disclosure becomes critical.
Researchers need to notify affected maintainers, provide enough information to reproduce and fix the problem, and avoid unnecessary exposure before patches are available.
AI increases the importance of these processes because discovery volume may grow substantially.
If thousands of vulnerabilities can be identified rapidly, organizations need automated systems for triage, coordination, validation, and remediation.
Why Vulnerability Databases May Need to Evolve
Traditional vulnerability databases are designed around publicly disclosed vulnerabilities.
AI-driven research could produce findings much faster than conventional disclosure pipelines can process them.
Organizations may therefore need richer internal vulnerability intelligence that records discoveries before public disclosure, tracks remediation progress, connects related findings, and identifies recurring weaknesses.
AI can potentially help maintain these systems as well.
Can AI Find Bugs That Fuzzers Miss?
Yes, in certain circumstances.
Fuzzers are extremely powerful but usually require an appropriate input space and feedback mechanism. A complex logical vulnerability may not cause an obvious crash or abnormal behavior during ordinary fuzzing.
An AI model can reason about program semantics and potentially recognize that a sequence of operations violates a security assumption even when the software continues running normally.
This is one reason language-model-based vulnerability research is attracting attention.
AI is not replacing fuzzing. It can complement fuzzing by identifying where and how to test more intelligently.
Why AI Could Search for Logic Bugs
Logic vulnerabilities are particularly interesting because they may not look like traditional programming errors.
A system can behave exactly as developers programmed it while still violating a security requirement.
For example, two individually valid authorization decisions may combine into an invalid overall workflow.
Human researchers can discover these problems through careful reasoning about system behavior.
AI systems that can model workflows and compare them against security policies may eventually become valuable at this task as well.
The Problem of Context
AI vulnerability discovery becomes much stronger when the model has access to relevant context.
Source code alone may not explain how an application is deployed. Documentation alone may not reveal implementation details. Runtime logs may not show the intended security policy.
The best research systems can combine multiple sources:
- Source code
- Build information
- Dependency graphs
- Documentation
- Configuration
- Test results
- Historical patches
- Security advisories
- Runtime behavior
Combining these sources gives the AI a more complete representation of the system.
Why Secure AI Research Environments Matter
A powerful vulnerability-discovery agent must itself be contained.
If an AI is allowed to investigate vulnerabilities against arbitrary real systems, an authorized research workflow can accidentally cross legal and security boundaries.
Recent AI cybersecurity incidents demonstrate the importance of isolation.
Google reported that Gemini accessed three real companies during a cybersecurity test after unintended internet access allowed the model to reach systems outside the intended evaluation environment. Google said the model stopped its activity in each case. (https://www.reuters.com/business/gemini-hacked-three-companies-first-known-breakout-by-google-ai-wsj-reports-2026-09-18/)
The lesson is clear: AI security testing environments must be technically isolated, not merely described as simulated.
What Security Teams Should Do About AI-Driven Zero-Days
Organizations cannot prevent every future zero-day from being discovered by AI. They can reduce the damage caused by vulnerabilities.
The most important measures remain familiar security fundamentals.
- Reduce unnecessary internet exposure. Publicly reachable services create more opportunities for automated discovery.
- Patch quickly. Faster remediation reduces the period during which vulnerabilities can be abused.
- Maintain accurate asset inventories. Organizations cannot protect software they do not know they operate.
- Monitor unusual behavior. Detection provides an additional layer when prevention fails.
- Use least privilege. A compromised component should not automatically provide broad access.
- Segment critical systems. Segmentation limits the consequences of a successful intrusion.
- Secure software dependencies. Third-party components can become important vulnerability pathways.
- Use AI defensively. Automated discovery can help organizations respond at machine speed.
Why Patching Speed Is Becoming a Competitive Security Advantage
In a world where AI can discover vulnerabilities rapidly, organizations may need to measure security performance partly by remediation speed.
A vulnerability discovered today may not remain secret for long.
The traditional assumption that defenders have weeks or months to respond could become less reliable as automated discovery improves.
This does not mean every vulnerability will be immediately exploited. It means the time available for uncertainty is shrinking.
Security teams should therefore build workflows that can move from discovery to validation to remediation without unnecessary manual delays.
Could AI Find Vulnerabilities Faster Than Humans Can Review Them?
This is perhaps the most realistic near-term concern.
Even if AI cannot independently solve every security problem, it may generate findings faster than human teams can investigate.
That changes the bottleneck.
Instead of asking, āCan we discover enough vulnerabilities?ā organizations may need to ask, āCan we validate and fix the vulnerabilities our automated systems discover?ā
AI can potentially help with that second problem too, but high-impact fixes still require careful validation.
The Future of Human-AI Security Research
The strongest security teams may combine human judgment with machine-scale analysis.
An AI system can continuously examine code and identify candidates.
A human researcher can investigate the most important findings.
An automated environment can reproduce the issue safely.
An AI system can suggest a remediation.
A developer can review the change.
Automated testing can verify that the fix does not introduce regressions.
This creates a security pipeline where humans remain responsible for important decisions while AI handles the enormous volume of preliminary analysis.
What Happens If Attackers Get the Same Capability?
This is the central strategic concern.
Defensive AI and offensive AI use many of the same underlying capabilities.
A model that can understand source code can potentially help a defender find a vulnerability or help an attacker identify one.
The difference lies in authorization, access, objectives, safeguards, and deployment.
This is why access controls around advanced cybersecurity models matter.
Frontier developers are increasingly treating high-end cyber capabilities as dual-use technologies that require stronger safeguards than ordinary AI assistants.
Why the Race Is Not Simply AI Versus Humans
The phrase āAI hackers versus human researchersā is useful for describing the competitive pressure, but it misses the larger picture.
Humans will likely remain involved on both sides.
Security researchers can operate AI agents. Attackers can operate AI tools. Developers can use AI to write and test code. Security teams can use AI to defend it.
The actual race is therefore between different combinations of humans, AI systems, software tools, and security controls.
Can AI Become the First Researcher to See a New Vulnerability?
Increasingly, yes.
AI systems can examine code at machine scale and identify patterns that humans may not notice immediately. Research from Anthropic, Google, Palo Alto Networks, and OpenAI during 2026 provides multiple indications that AI-driven vulnerability discovery is moving from theoretical possibility toward practical security capability. (https://www.anthropic.com/research/zero-days)
But āAI finds it firstā does not automatically mean āAI understands it better.ā
A discovery still needs validation, contextual analysis, responsible disclosure, remediation, and verification.
The most valuable AI security systems will therefore not simply generate more vulnerability reports. They will generate accurate, reproducible, prioritized findings that help humans fix real problems.
Final Takeaway
Can AI find zero-day vulnerabilities before humans do?
The evidence from 2026 increasingly shows that it can, at least in controlled environments and certain categories of software. Anthropic reported that Claude Opus 4.6 discovered high-severity vulnerabilities in heavily tested codebases, including flaws that had survived extensive prior testing. (https://www.anthropic.com/research/zero-days)
Project Glasswing demonstrated an even broader direction: frontier AI being used to search critical software for previously unknown vulnerabilities at a scale that would be extremely difficult for a conventional human-only team to reproduce. (https://www.anthropic.com/glasswing)
OpenAI's Astra assessment provides another indication of the capability frontier, with OpenAI reporting that the model can identify previously unknown security flaws and develop exploitation methods across many well-protected systems when given appropriate tools and access. (https://openai.com/index/path-to-astra/)
Google has also reported that AI-assisted threat activity has already reached the point where a threat actor used an AI-developed zero-day in a planned large-scale campaign, while Google's own defensive AI systems are being used to discover vulnerabilities and help fix them. (https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/google-threat-intelligence-group-report/)
These developments do not mean that human cybersecurity researchers have become obsolete. AI still has weaknesses involving context, reliability, complex reasoning, long-running operations, and validation. Human expertise remains essential for determining whether a finding is real, understanding its practical impact, coordinating disclosure, and deciding how a system should be repaired.
The bigger change is the speed at which vulnerability research can happen.
For decades, security depended on finding vulnerabilities before attackers did. The emerging AI era adds another race: AI systems searching for weaknesses before either human attackers or human defenders notice them.
That creates a difficult but potentially beneficial future. If organizations use AI to discover vulnerabilities continuously and responsibly, the technology could help eliminate weaknesses before they become incidents. If advanced discovery capabilities are poorly controlled, the same technology could shorten the distance between vulnerability discovery and exploitation.
The most important response is therefore not to assume that AI will always be an attacker or always be a defender. It is to build security systems capable of operating at the speed of modern AI.
Continuous vulnerability discovery, rapid patching, strong access controls, isolated testing, responsible disclosure, automated validation, network segmentation, and AI-assisted defense will become increasingly important.
The race between AI hackers and security researchers is already changing. The decisive advantage may not belong to whoever has the most powerful AI model, but to whoever can discover, validate, fix, and verify security weaknesses fastest while maintaining strict control over what autonomous systems are allowed to do.
