The short answer is: not because they are “evil,” but because they dramatically expand what a capable AI system can do—especially in domains like cybersecurity. The danger comes from capability plus scale, not from malicious intent.
Anthropic itself describes Fable 5 as the public version of a new “Mythos-class” model, while Mythos 5 is essentially the same underlying model with some safety classifiers removed for trusted partners.
The concerns fall into several categories.
1. Cybersecurity capability
This is the biggest reason.
Anthropic has stated that Fable 5 is extremely capable at software engineering and long-horizon coding tasks. To reduce misuse, requests involving offensive cybersecurity are automatically redirected to a less capable model (Opus 4.8). Mythos 5 does not have the same public-facing restrictions and is reserved for approved organizations.
The implication is that the model can:
- analyze millions of lines of code
- discover subtle vulnerabilities
- chain together attack paths
- generate sophisticated exploits
- continue working for hours or days
Those same abilities are enormously valuable for defense, but they also lower the cost of offensive cyber operations if misused.
2. Long-horizon agency
Older LLMs usually worked one prompt at a time.
Fable 5 was designed for long-running agentic work.
According to Anthropic, it can
- plan over many stages,
- delegate subtasks,
- verify its own work,
- continue operating for extended periods with little human intervention.
That means the question is no longer
“Can the AI write code?”
but rather
“Can the AI carry out an entire software engineering project almost autonomously?”
That represents a qualitative change.
3. Dual-use technology
Nearly every powerful capability has two sides.
The same model can
- secure hospitals,
- protect power grids,
- audit aircraft software,
or
- automate vulnerability discovery,
- accelerate malware development,
- help sophisticated attackers.
This is why governments increasingly describe frontier AI as a dual-use technology, analogous to advanced cryptography or satellite imaging.
4. Why did the U.S. government intervene?
This month, the U.S. government temporarily restricted access to Mythos 5 and Fable 5 over national security concerns before later allowing limited access to selected trusted organizations after negotiations with Anthropic. The stated concerns centered on advanced cyber capabilities and potential foreign misuse.
That does not necessarily mean the models were uncontrollable.
Rather, policymakers judged that
the capability frontier had advanced enough to justify temporary restrictions while new governance mechanisms were developed.
5. The “danger” is really a capability threshold
I think the deeper issue is this.
For many years we asked
“Can AI answer questions?”
Now we ask
“Can AI accomplish objectives?”
That is a different category.
An AI that can pursue goals over hours or days becomes much more economically—and strategically—important.
6. Your earlier theme: the basin of attraction
Since we’ve recently been discussing strange attractors and attractor basins, I think there is an interesting connection.
Models like Mythos 5 are dangerous not merely because they are “smarter.”
They are dangerous because they reshape the attractor landscape of civilization.
Before:
Human expertise was the scarce resource.
After:
Autonomous cognitive labor becomes abundant.
Once that happens,
- software development,
- scientific research,
- cyber defense,
- intelligence analysis,
- engineering,
all begin to flow toward a new basin of attraction.
In that sense, Mythos 5 is less like “a better chatbot” and more like the emergence of a new cognitive infrastructure. The societal risks arise not only from what the model can do, but from how it changes the distribution of expertise, power, and decision-making.
That, in my view, is the most profound reason frontier models such as Fable 5 and Mythos 5 are being treated differently from previous generations of AI.
