We put $5.5M behind a future worth living in. Awarded live on September 25. (Sponsor)

The futures we picture are the ones we end up building. So we put $5.5M behind a good one.

On September 25 in downtown LA, ten filmmakers out of nearly 20,000 present a future worth living in, and one leaves with $2.5M to go make it real. The same night, $2M goes to founders who used AI to solve a problem people actually have.

1,500 seats. Hosted by Peter Diamandis, with Google, Circle and XPRIZE. General admission $1,495.

Claim your seat

The most interesting thing about Anthropic’s latest cybersecurity move is not that Claude Mythos 5 can scan code for vulnerabilities.

It is the way Anthropic is choosing to expose that capability.

Instead of giving every security engineer a prompt box connected directly to its most capable cybersecurity model, Anthropic is increasingly putting Mythos behind specialized interfaces. The model can inspect a repository, reason about vulnerabilities and generate remediation advice—but the user receives the result of that reasoning rather than unrestricted access to the reasoning engine itself.

In Claude Security, for example, Mythos 5 can return a vulnerability finding with a CWE category, confidence rating, severity assessment and suggested fix. The user can then move into Claude Code to implement the remediation using models already available to their organization. Mythos itself does not suddenly become available as a general-purpose interactive model, and fixes require human approval.

That distinction may look like a product-design detail.

It is actually a potentially important security architecture for frontier AI.

The cybersecurity capability problem is becoming different

For most of the LLM era, the cybersecurity debate centered on whether models could help attackers write phishing emails, generate scripts or explain known vulnerabilities.

Those capabilities mattered, but much of the output still required a knowledgeable operator.

Frontier cyber models are moving into a different category.

Anthropic says its earlier Mythos Preview model demonstrated the ability to autonomously discover vulnerabilities, construct working exploits and combine multiple vulnerability primitives into attack chains. In testing, it chained two, three and sometimes four Linux vulnerabilities together; Anthropic also reported successful vulnerability exploitation across major web browsers and other complex targets.

The important change is not simply that the model “knows cybersecurity.”

It is that the model can increasingly perform work previously requiring experienced vulnerability researchers.

That collapses several traditional barriers to offensive security:

• Expertise barrier - A sophisticated exploit might previously require someone who understood memory corruption, kernel behavior, browser internals or application logic. AI increasingly supplies part of that expertise.

• Time barrier - Tasks taking human researchers days or weeks can potentially be parallelized across many model instances.

• Scale barrier - One researcher can investigate a handful of systems, while thousands of agents can investigate thousands.

• Iteration barrier - Exploitation often involves repetitive debugging: attempt something, inspect the crash, modify the approach and repeat. Agents are unusually well suited to this kind of loop.

The danger therefore comes less from any individual exploit technique and more from the potential industrialization of vulnerability research.

Discovery and exploitation are converging

Historically, vulnerability discovery and exploitation were somewhat distinct disciplines.

A scanner might tell you:

Something suspicious exists here.

A security researcher then needed to determine whether it was actually exploitable.

Frontier models increasingly narrow that gap.

Conceptually, the workflow is moving toward:

SOURCE CODE
     ↓
Understand architecture
     ↓
Trace trust boundaries
     ↓
Identify vulnerability
     ↓
Validate vulnerability
     ↓
Construct exploit primitive
     ↓
Chain primitives
     ↓
Test exploit
     ↓
Operational attack

The farther a model can travel down that chain autonomously, the more consequential unrestricted access becomes.

Anthropic’s own cybersecurity research highlights this issue. Mythos Preview showed substantial improvements not only at identifying vulnerabilities but at transforming them into exploit primitives and combining those primitives into functioning attack chains.

That creates a difficult problem for AI companies.

The same reasoning ability that makes the model extremely valuable to defenders is almost exactly what makes it valuable to attackers.

The traditional answer would be: restrict the model

There is an obvious solution.

Do not release it.

And Anthropic has indeed kept direct Mythos access limited. Mythos 5 remains positioned as a model for trusted cybersecurity and research access rather than normal unrestricted use. Its broadly available counterpart, Claude Fable 5, uses the same underlying model family while applying safeguards around sensitive cybersecurity and biological capabilities.

But simply withholding frontier models introduces another problem.

Attackers do not need Anthropic specifically.

Anthropic argues that comparable cyber capability is likely to emerge across the broader model ecosystem.

If powerful offensive capability eventually becomes available somewhere—through another provider, stolen weights, open-weight models, jailbreaks or future systems—then permanently preventing defenders from using frontier AI leaves them at a disadvant