AI So Dangerous They Locked It Away
Updated: 1 day ago
ANTHROPIC · AI SECURITY · 2026 Update
Claude Mythos found 27-year-old security flaws overnight that every scanner on earth had missed. So instead of shipping it, Anthropic handed it to Apple, Microsoft, Google, Amazon and told the rest of us to wait patiently ...
AI So Dangerous They Locked It Away
There is a version of this story where Anthropic is the hero.
And there is a version where the whole thing is the most elegant piece of corporate theatre the tech industry has ever produced.
The truth, as usual, is probably somewhere in between and far more interesting than either camp wants to admit.
On April 7, 2026, Anthropic quietly announced Claude Mythos Preview. A model so capable at hunting security vulnerabilities that they decided the general public simply couldn't be trusted with it yet. Engineers set it loose on codebases overnight and woke to thousands of bugs found, categorised, and already equipped with working exploits.
Done. Ready to go. In a single night.
"It isn't AI that writes code anymore. It's AI that breaks it and that distinction changes everything about how we think about digital security."
What It Actually Found
The numbers are staggering. In the weeks since its controlled release, Mythos Preview has surfaced thousands of zero-day vulnerabilities across every major operating system and web browser. But the ages of some of these bugs are what should really unsettle you.
These were not obscure edge cases. These were bugs that had survived millions of scans by automated tools, decades of peer review, and some of the best security engineers on the planet.
Mythos found them overnight because it doesn't scan code the way traditional tools do, it reasons about it, the way a deeply experienced human attacker would.
Except faster. And it never sleeps.
The Alliance They Built Around It
AI So Dangerous They Locked It Away
Rather than releasing Mythos to the world, Anthropic launched Project Glasswing named for a butterfly whose transparent wings allow it to hide in plain sight, much like the vulnerabilities the model is designed to expose.
The initiative brings together twelve founding partners:
Amazon Web Services
Apple, Broadcom
Cisco
CrowdStrike
Google
JPMorganChase
The Linux Foundation
Microsoft
NVIDIA
Palo Alto Networks
Anthropic itself.
Over 40 additional organisations building critical internet infrastructure were also quietly handed access.
The stated goal: give defenders a head start. Use Mythos to patch the most important systems on the internet before its capabilities become broadly available, either through a public release or through a rival building something similar
The window they're working in is narrow. Anthropic has privately warned senior government officials that models with these capabilities make large-scale cyberattacks significantly more likely within the year.
The implication is clear: they're in a race against themselves & against whoever comes next.
THE QUESTIONS NOBODY IS ASKING LOUDLY ENOUGH
Here is where it gets complicated. Because however responsible the Project Glasswing initiative looks on paper, the structure of it raises questions that deserve serious scrutiny.
The companies given early access are also the companies with the most to gain from a more secure internet and the most to lose from a rival having this capability first.
Apple, Microsoft, Google, and Amazon are not neutral parties.
Handing the world's most powerful security AI exclusively to the world's most powerful technology companies is not the same thing as protecting the internet. It is protecting the incumbents.
Smaller software vendors, open-source maintainers, and developers in emerging markets the people who arguably run the most vulnerable parts of the internet's infrastructure are not in the room.
Anthropic has pledged $4M in donations to open-source security organisations, which is a start. But against the scale of what Mythos can do, it reads more like an afterthought than a strategy.
"The most unsettling part isn't the AI itself. It's the realisation that the decisions about who gets to use it were made entirely behind closed doors."
THE HITL DEBATE AND WHY IT'S ALREADY OUTDATED
AI So Dangerous They Locked It Away
For years, the cybersecurity industry has organised itself around a reassuring principle:
Humans stay in the loop.
Every alert gets reviewed.
Every action gets approved.
A machine might flag a threat, but a person decides what to do about it.
That principle is now under serious strain and Mythos has accelerated the argument to breaking point in some industries.
The problem here is one of scale. For example large financial institutions now generate over 200,000 security alerts per day.
Analysts investigate fewer than 5% of them. The human loop isn't a safeguard at that volume it's a bottleneck with a false sense of security stapled to it.
The shift from HITL to HOTL isn't just a workflow change it's a transfer of liability.
When a human approves every action, the chain of accountability is clear.
When an AI operates autonomously within guardrails and those guardrails fail, the legal and ethical frameworks to deal with the consequences simply don't exist yet. Vodafone's global CISO, Emma Smith, made exactly this argument at RSAC 2026, calling for the industry to accept that "human on the loop" is not a concession it's an evolution.
Not everyone agreed. The debate is very much live.
"The human in the loop was never really a safety feature. At 200,000 alerts a day, it was theatre. The question now is what genuine oversight actually looks like when the machine moves faster than any team can follow."
What makes Mythos different is that it doesn't just trigger alerts for humans to review it reasons about them, chains them together, develops exploit paths, and writes working code. Putting a human in that loop isn't reviewing a flag.
It's asking a security analyst to validate the output of something operating at a cognitive level they may not be able to independently verify. That's a very different kind of oversight. And right now, the industry doesn't have a consensus answer for how to handle it.
THE GOVERNANCE GAP NOBODY IS TALKING ABOUT
Underneath all of this sits a regulatory reality that is quietly alarming.
The World Economic Forum's Cybersecurity Outlook 2026 found that 87% of security leaders now identify AI vulnerabilities as the fastest growing risk in their environment.
IBM's 2025 data shows that only 37% of organisations have AI governance policies in place.
The remaining 63% are operating without guardrails and many of them are running AI agents they didn't deliberately deploy.
Microsoft's Cyber Pulse 2026 found that 29% of employees are already using unsanctioned AI agents for work tasks.
This is the shadow infrastructure that Mythos will eventually be turned loose on.
Not the carefully patched systems of Apple and Microsoft, but the thousands of mid-sized enterprises running unmonitored AI agents in environments with no governance framework, no visibility into what those agents are doing, and no audit trail when something goes wrong.
Anthropic's Project Glasswing addresses the top of the stack, the most critical, most watched, most resourced systems on earth.
It does not address everything below it. And everything below it is where most of the actual damage from a broad capability release would land.
That is the conversation the industry needs to be having. Not just about who got early access to Mythos, but about what the world looks like for the organisations that didn't and what it will cost them when the window finally closes.
TURNING GOVERNANCE INTO A COMPETITIVE ADVANTAGE
The message here is not to slow down AI adoption.
It’s the opposite.
The organisations that will benefit most from AI over the next 12–24 months are not the ones avoiding it they’re the ones using it with control, clarity, and confidence.
Good governance is what makes that possible.
AI is no longer a single category of tooling.
It spans everything from productivity assistants and customer support bots to code generation, security automation, HR screening, marketing optimisation, and autonomous decision systems. Many of these tools operate quietly in the background of day-to-day business activity often introduced at team level, outside formal procurement processes, and without full visibility of how data is being processed.
That is exactly where risk begins to accumulate.
From a data protection and GDPR perspective, the question is not whether you should be using AI.
It is whether you understand:
What data is being input,
How that data is being processed,
Where it is stored or transferred,
What decisions are being made as a result.
This applies across a wide range of common business use cases, including:
AI-powered customer service platforms handling personal data at scale
HR and recruitment tools screening candidates or analysing employee performance
Marketing and analytics platforms profiling user behaviour and generating targeted content
Developer tools generating or reviewing code with access to proprietary systems
Security tools autonomously identifying vulnerabilities or responding to threats
Document processing systems extracting and classifying sensitive information
Each of these can fall directly within GDPR scope particularly where personal data, automated decision-making, or profiling is involved.
The organisations that get this right are not the ones with the longest policies.
They are the ones that can demonstrate control:
Clear AI system inventories and risk classifications
Data flow mapping across internal and third-party tools
Embedded Data Protection Impact Assessments
Defined human oversight for high-risk decisions
Tested audit trails and incident response procedures
Vendor governance that treats third-party risk as first-party exposure
This is where governance shifts from compliance burden to operational strength.
For a free no obligation quote on how CompliancePigeon can support your business with data protection and cyber security compliance and guidance please contact us.
We have a range of services from data protection auditing and fractional DPO support to pen testing and cyber security implementation.
Comments