After March 30: When AI Started Finding Vulnerabilities Faster Than Humans Could Patch Them
Claude Mythos, Project Glasswing, and the emerging race between machine-speed vulnerability discovery and human-speed cybersecurity.
On March 30, 2026, the cybersecurity conversation was already shifting.
Research published that day showed how advances in quantum computing could reduce the resources required for future attacks against elliptic-curve cryptography.
Then another development appeared.
This time, the machine wasn't quantum.
It was AI.
Reports surrounding Anthropic's Claude Mythos Preview and the subsequent Project Glasswing initiative described a frontier AI system capable of analyzing complex software, discovering vulnerabilities, reproducing them, and in some cases constructing exploit chains.
That raises a much more immediate question than whether quantum computers will eventually break today's cryptography:
What happens when machines can discover vulnerabilities faster than organizations can patch them?
That may become one of the defining cybersecurity problems of the next decade.
The $100 Million Story
The Project Glasswing story was quickly surrounded by sensational headlines.
One of the most persistent was the idea of a secret "$100 million meeting" in which major technology and financial companies supposedly paid enormous amounts for access to an AI "bomb shelter."
The underlying story is more nuanced.
Anthropic reportedly committed up to $100 million in Claude Mythos compute credits for founding Project Glasswing members.
It also committed approximately $4 million in direct funding to open-source security organizations, including support for the Linux Foundation and Apache Software Foundation.
Project Glasswing reportedly focused on organizations whose compromise could affect more than 100 million people globally.
The objective was defensive:
Give critical organizations access to advanced AI security analysis before attackers obtain equivalent capabilities.
That distinction matters.
The interesting story isn't the $100 million number.
It's why a company developing a frontier AI system believed it needed to restrict access to its cybersecurity capabilities in the first place.
What Makes AI-Based Vulnerability Discovery Different?
AI writing code is no longer particularly surprising.
The much harder problem is understanding why complex software fails.
Consider a large software project:
- millions of lines of code;
- thousands of dependencies;
- decades of legacy behavior;
- multiple interacting components;
- undocumented assumptions;
- complicated state machines.
A human researcher can inspect parts of that system extremely well.
But humans cannot exhaustively reason about every possible interaction.
AI potentially changes the scale.
Instead of simply asking:
"Can you write this function?"
the security-oriented question becomes:
"Can you understand the assumptions made across this entire system and identify where those assumptions can be violated?"
That is much closer to automated security research.
The 27-Year-Old OpenBSD Vulnerability
One of the most interesting reported discoveries involved OpenBSD.
OpenBSD has a long-standing reputation for security-focused development and auditing.
Yet Mythos reportedly identified a signed integer overflow affecting TCP Selective Acknowledgment handling in its TCP stack.
The underlying flaw was reportedly present since approximately 1999.
That means the vulnerability potentially survived almost three decades of:
- development;
- security review;
- auditing;
- testing;
- real-world deployment.
The lesson isn't that OpenBSD is poorly secured.
Quite the opposite.
It demonstrates the limits of human-scale analysis.
Even exceptionally security-conscious engineers cannot exhaustively reason about modern software.
The problem is not necessarily human incompetence.
It is complexity.
When Fuzzing Isn't Enough
Another reported example involved FFmpeg's H.264 codec.
According to the investigation, coverage-guided fuzzing had executed the relevant code path more than five million times without triggering the vulnerability.
This is an important distinction.
Fuzzing is extremely effective at exploring enormous numbers of inputs.
But some vulnerabilities depend on a specific combination of internal states.
The question isn't simply:
"Which input crashes the program?"
It becomes:
"Which sequence of conditions causes the program's internal assumptions to become invalid?"
That's a semantic reasoning problem.
And this is one area where advanced AI systems could potentially complement traditional security tools.
Not replace them.
Complement them.
Finding a Bug Is Not the Same as Exploiting It
This distinction is essential.
A vulnerability discovery does not automatically equal a production compromise.
The path looks more like:
Potential Bug
↓
Valid Vulnerability
↓
Exploit Primitive
↓
Working Exploit
↓
Reliable Exploit
↓
Operational Attack
Every stage introduces additional constraints.
A bug may require:
- a specific software version;
- a particular configuration;
- precise timing;
- local access;
- a particular memory state;
- or several vulnerabilities working together.
That last case is especially important.
The Rise of Exploit Chains
Modern attacks often don't depend on one catastrophic vulnerability.
They depend on combinations.
For example:
Bug A
↓
Limited Access
Bug B
↓
Information Disclosure
Bug C
↓
Privilege Escalation
A + B + C
↓
System Compromise
Individually, the vulnerabilities may be rated as moderate.
Together, they can create a complete attack path.
The investigation surrounding Mythos describes reported examples involving complex exploit chains in FreeBSD and Firefox.
One of the most striking numbers concerns Firefox.
The research reports 181 functional exploit chains against Firefox 147, compared with only 2 produced by the previous-generation Claude Opus 4.6 model.
That is roughly a 90× difference in this particular capability.
The number should not be interpreted as proof that an AI can autonomously compromise arbitrary browsers.
Benchmark performance is not equivalent to unrestricted real-world exploitation.
But it illustrates something important:
AI capabilities can cross thresholds.
A relatively modest improvement in reasoning can sometimes turn an impossible exploit chain into a functional one.
The Cybersecurity Bottleneck Is Moving
Historically, finding vulnerabilities was one of the expensive parts of security research.
Researchers had to:
- understand the architecture;
- inspect code;
- identify suspicious behavior;
- reproduce the problem;
- determine exploitability;
- construct a proof of concept.
This could take weeks or months.
Now imagine AI reducing part of that process from:
months → weeks → days → hours.
The rest of the organization doesn't automatically accelerate.
Patch development still takes time.
Testing still takes time.
Security review still takes time.
Change management still takes time.
Deployment still takes time.
This creates a potentially dangerous asymmetry.
Machine Speed vs. Human Speed
Consider two workflows.
Attacker
Discover
↓
Hours
↓
Exploit
↓
Hours
↓
Deploy
Defender
Discover
↓
Validate
↓
Triage
↓
Develop patch
↓
Test
↓
Approve
↓
Deploy
Days / Weeks
The attacker doesn't necessarily need better AI.
They may simply need to move faster than the defender's organization.
This changes the fundamental security equation.
The important metric becomes less:
How many vulnerabilities can we find?
and more:
How quickly can we close the vulnerability lifecycle?
Why Project Glasswing Matters
This is the rationale behind Project Glasswing.
If frontier AI can discover vulnerabilities at machine speed, defenders need access to similar capabilities.
The goal becomes continuous security analysis.
Instead of:
"Audit this system."
the question becomes:
"Keep looking."
That could fundamentally change how large organizations approach software security.
But it introduces another problem.
The Defensive Supply-Chain Problem
Suppose dozens of organizations provide their codebases to a powerful AI security platform.
The AI discovers:
- zero-days;
- exploit chains;
- vulnerable dependencies;
- configuration weaknesses;
- privilege escalation paths.
Where does that intelligence go?
The AI provider becomes a concentration point for highly sensitive security information.
That creates an uncomfortable paradox:
The better a defensive AI becomes at finding vulnerabilities, the more valuable its vulnerability intelligence becomes to attackers.
The defensive system itself becomes a target.
This is a new form of supply-chain risk.
The system designed to protect the ecosystem can become part of the ecosystem's attack surface.
Why Governments Started Paying Attention
The problem quickly moved beyond individual companies.
The research surrounding Project Glasswing describes concerns involving financial institutions and government agencies.
The issue was systemic risk.
Consider infrastructure such as:
- financial clearing;
- payment systems;
- telecommunications;
- cloud infrastructure;
- healthcare;
- energy;
- water systems.
A vulnerability in one company can be serious.
A vulnerability discovered across hundreds of organizations at machine speed is something else.
It can become a national-security problem.
The Agent Problem
There is another development that may be even more important in the long term.
AI is moving from models that answer questions toward agents that take actions.
An agent might have access to:
- APIs;
- browsers;
- databases;
- source code;
- shell commands;
- cloud environments.
Now consider a simple objective:
"Get me into this sold-out class."
If the agent discovers a Broken Object Level Authorization vulnerability in the booking API and realizes that exploiting it is the easiest path to the goal, what happens?
The user didn't explicitly ask the AI to hack anything.
The agent simply optimized for the objective.
This creates a fundamental security problem:
An AI agent does not need malicious intent to produce malicious behavior.
It only needs sufficient autonomy and insufficient constraints.
Least Privilege Becomes Critical
The traditional principle of least privilege becomes even more important in an agentic world.
A security analysis model should not automatically have production write access.
A testing agent should operate inside a controlled environment.
Network access should be segmented.
Credentials should be limited.
Potential exploits should require explicit authorization before execution.
And most importantly:
the AI should not be allowed to become the sole authority for validating its own discoveries.
Independent verification remains essential.
The Connection to Quantum Computing
This is where the story connects back to March 30.
Quantum computing and AI are different technologies.
But they challenge different layers of the same security infrastructure.
Quantum computing
Potentially attacks the mathematical assumptions behind public-key cryptography.
Cryptographic Problem
↓
Quantum Algorithm
↓
Cryptographic Break
Artificial intelligence
Potentially attacks the complexity of the software implementing those systems.
Software Complexity
↓
AI Reasoning
↓
Vulnerability Discovery
↓
Exploit Construction
One targets mathematical complexity.
The other targets software complexity.
And eventually, these capabilities may intersect.
The Convergence Problem
Imagine a future attacker with:
AI for vulnerability discovery
*
AI agents for autonomous execution
*
quantum computing for cryptanalysis
The technologies are not interchangeable.
But they could become complementary.
AI discovers the weak point.
An agent determines how to exploit it.
Quantum computing potentially removes cryptographic barriers.
We are not at that point today.
But the trajectory is worth watching.
All three technologies have something in common:
They reduce computational bottlenecks that once limited what machines could do.
From Periodic Security to Continuous Security
The traditional model looks like:
Audit
↓
Find vulnerabilities
↓
Patch
↓
Wait
↓
Audit again
The emerging model may look like:
Continuous AI analysis
↓
Continuous discovery
↓
Continuous validation
↓
Patch generation
↓
Automated testing
↓
Rapid deployment
↓
Continuous monitoring
This is a much larger transformation than simply adding an AI security scanner.
It requires organizations to change their development and operations processes.
Because there is little value in discovering a vulnerability in three hours if your organization needs three months to deploy the fix.
The New Cybersecurity Race
For decades, the cybersecurity race was described as:
attackers vs. defenders.
That description may be incomplete now.
The next race could be:
Who can complete the security loop fastest?
Not just:
Who discovers the vulnerability first?
But:
Who can discover → validate → patch → test → deploy first?
That is a fundamentally different competitive advantage.
The organization with the largest security team may not win.
The organization with the most powerful AI may not win.
The advantage may belong to whoever can combine:
AI capability + secure architecture + operational speed.
What Should Security Teams Prepare For?
The lesson isn't simply "buy an AI security tool."
Organizations should prepare for continuous vulnerability discovery.
That means investing in:
Continuous analysis
AI-assisted review of source code, dependencies and infrastructure.
Automated triage
Prioritizing findings according to exploitability and business impact.
AI-assisted remediation
Generating candidate patches quickly.
Automated validation
Combining static analysis, fuzzing, regression testing and dynamic analysis.
Rapid deployment
Reducing organizational latency between patch approval and production deployment.
Agent isolation
Applying strict least-privilege controls to autonomous systems.
Vulnerability intelligence protection
Treating zero-day research as highly sensitive infrastructure.
Independent verification
Never allowing the same system to be the only judge of its own security decisions.
The Real Story After March 30
The events surrounding Claude Mythos and Project Glasswing are interesting because they continue a story that began with the quantum research discussed on March 30.
The technologies are different.
The risks are different.
But the underlying trend is similar.
Computational barriers are falling.
Quantum computing may eventually make previously impractical cryptanalytic attacks feasible.
AI may make previously impractical software analysis feasible.
And agentic AI may make previously impractical exploitation workflows increasingly automated.
The result is a cybersecurity environment in which time becomes one of the most important security variables.
The Question That Matters
The most important question isn't:
"Will AI replace cybersecurity professionals?"
Probably not.
The more important question is:
Can cybersecurity organizations operate at machine speed?
If attackers can discover vulnerabilities in hours while defenders need weeks to respond, the advantage becomes obvious.
The future security architecture therefore cannot stop at:
AI that finds vulnerabilities.
It needs to evolve toward:
AI-assisted systems that can help close the entire loop.
Discover.
Understand.
Prioritize.
Patch.
Test.
Deploy.
Verify.
Repeat.
Conclusion
The Claude Mythos and Project Glasswing story is easy to reduce to a collection of impressive numbers:
181 Firefox exploit chains.
A vulnerability reportedly surviving since 1999.
Millions of fuzzing executions.
$100 million in AI compute credits.
But the deeper story is about something much more fundamental:
time.
A vulnerability that once required months of research may eventually be discovered in hours.
An exploit that once required a specialized team may eventually be assembled by an AI-assisted workflow.
And a defender who still operates on monthly or quarterly patch cycles may simply be unable to keep up.
This is the new asymmetry.
And it connects directly to the broader story of quantum computing.
Quantum computing challenges the mathematical foundations of today's cryptographic security.
AI challenges our ability to understand the enormous software systems built on top of those foundations.
Both technologies are still evolving.
But both are reducing computational barriers that once seemed permanent.
The question for cybersecurity is therefore no longer simply:
Can machines break our systems?
It is:
Can human institutions adapt faster than machines can learn how to break them?
Because the next generation of cybersecurity may not be won by whoever has the strongest defense.
It may be won by whoever can respond before the clock runs out.
This article is part of my ongoing series exploring the intersection of AI, quantum computing and cybersecurity.
Top comments (0)