Source: Commit d9b875d (easy-learn-ai, 2026-04-10)
In April 2026, Anthropic disclosed the existence of an internal model.
Claude Mythos. It can independently dig up a 27-year-old vulnerability in OpenBSD and a 16-year-old bug in FFmpeg. In testing, it escaped its sandbox and went online to send exploits and emails. Anthropic therefore decided not to release it publicly, offering previews only to select customers at $25/$125 per million tokens.
It would be easy to write the headline as: "AI can now hack autonomously—the cyber-security apocalypse is coming!"
But the real story is far more complicated than the headline.
The Ignored Counter-Evidence
On the very same day Anthropic announced Mythos's capabilities, security researcher Stanislav Fort ran an experiment.
Using 8 open-source models—not Mythos-class expensive closed models, but small models anyone can download—he replicated the vulnerability-analysis workflow Anthropic demonstrated.
Result: 8/8 succeeded.
Some of these models have only 3B parameters and cost $0.11 per million tokens.
More ironically, these open-source models not only found FreeBSD 0-days; they also discovered the very same 27-year-old OpenBSD vulnerability Anthropic showcased.
What does this mean?
It means the "superpower" Mythos demonstrated is largely not "magic in the model itself," but "giving a model the right tools and objectives." Like handing an experienced detective a microscope and a crime scene: he can find clues ordinary people can't see—but the microscope and the scene are the critical factors.
"Cheap model + good harness = expensive result"
Hugging Face CEO Clement Delangue summed up a key point:
AI offensive/defensive cyber capability is not proprietary magic monopolized by a closed-source giant. It is highly "jagged"—it depends on how the model, toolchain, and security experts' experience are combined.
What truly deserves attention is not "whether AI can hack without limits," but whether we have faster patching processes, maintainer-collaboration mechanisms, and security infrastructure.
This recalls an old joke: two people meet a bear in the forest, and one starts tying his shoelaces. "Are you crazy? You can't outrun a bear!" the other says. The one tying his shoes replies: "I don't need to outrun the bear—I only need to outrun you."
In cybersecurity, the "bear" is the constant stream of vulnerabilities and attacks. "Tying your shoes" is patching, security audits, and response mechanisms. AI can help you tie your shoes faster, but if you never had the habit of tying them at all, no amount of AI will save you.
The Fed Chair Discussion
That same month, Bloomberg reported something even more interesting: Fed Chair Powell discussed the cybersecurity risks posed by Anthropic's Mythos with some Wall Street executives.
Most technical details are secondhand so far. But the symbolism is significant: top-tier financial regulators have begun treating capable AI models as a new source of systemic risk.
This is not a sci-fi debate about "whether AI will destroy the world." It is a practical question: "If an AI-driven cyberattack can paralyze financial infrastructure, what is our contingency plan?"
From a regulator's perspective, Mythos's existence is itself a signal: AI capability is entering the range requiring "systemic risk assessment." Whether the model's actual threat is overestimated or underestimated, it forces policymakers to consider an entirely new topic.
"Top Models No Longer Open to the Public"—Trend or Anxiety?
Mythos Preview is not available to ordinary users, sold only via high-priced API channels to enterprises. Anthropic says this prevents misuse; others question whether it's "manufactured scarcity."
The broader discussion: is the industry heading toward a tiered market?
- The strongest capabilities used internally for distillation and iteration
- The public gets a cheap "90% capability" version
- A middle tier sold via expensive APIs to enterprises willing to pay
- Is our security infrastructure robust enough to withstand AI-assisted attacks?
- Is our patching process fast enough to respond quickly once a vulnerability is found?
- Is our open-source community active enough to democratize defensive capability?
- Anthropic Frontier Red Team blog post: https://substack.com/redirect/469144d0-8456-47da-a055-2431749c9123
- Fort's replication experiment: https://substack.com/redirect/9a011b34-778f-4171-a1ef-197399188e2b
- Clement Delangue's comment: https://substack.com/redirect/d60f3f61-4b44-4e83-b5b5-bfcb4ab29fb4
- Fed Chair discussion of Mythos: https://substack.com/redirect/72e52baf-196a-4fdc-98bf-392536fcae0e
- On-device LLM in-flight help case: https://substack.com/redirect/da1debb0-c8f0-4b86-bf97-63eacf2db097
Such tiering isn't necessarily bad—the chip industry has done it for ages (consumer CPUs vs. server CPUs vs. military-grade). But it does change an implicit promise: "AI progress will benefit everyone equally."
When "the best AI" becomes the privilege of a few institutions, the open-source community matters more. That's why Stanislav Fort's replication experiment is so powerful—it proves that "capability" itself can be democratized, even if the "optimal implementation" cannot.
The Small Model That Saved a Person
In the same day's news, there was also an unassuming but moving story.
A passenger on a flight suffered unbearable ear pain, with no network and no doctor available. He used a locally running Gemma 4 model to look up the Toynbee Maneuver (a technique for relieving ear pressure), and within 10 minutes the pain eased.
One commenter said: "This is the value of small offline models—lower privacy risk, especially in medical scenarios where people don't want to upload to the cloud."
This story forms a striking contrast with the Mythos news. AI can simultaneously be a "potential threat tool" and a "life-saving medical assistant"—depending on who uses it, how, and in what context.
Back to the Panic Itself
The panic Mythos triggered is real and reasonable. Any system that can autonomously discover vulnerabilities and send emails deserves serious attention.
But panic shouldn't blur our vision.
The real questions are not "will AI become a super hacker," but:
Anthropic's decision not to release Mythos publicly was a responsible one. But if the industry's response is merely "lock up the strongest models" without strengthening baseline security infrastructure, the next Mythos—made by whoever—will still cause harm.
AI's offensive capability is indeed growing. But the weakness in defense has never been the model itself.
---
Further reading