10io — Actual Intelligence | Fractional CAIO / Chief AI Officer

A Glomar Trap encompasses the type of factors that led companies in the 1970s to make investments to capitalize on the potential of deep-sea mining, the viability of which had been fabricated by the CIA as a cover story for the effort to secretly recover a sunken Soviet submarine. In other words, FOMO leading to YOLO at industry-scale.

Upon witnessing the reactions to the Claude Mythos Preview and related intentional and unintentional disclosures by Anthropic, the first thing that came to mind is that: Nothing that is claimed about Mythos should be surprising. The only thing that is surprising about the emerging Mythos mythos is that anyone is surprised by it. Mythos’ behavior is an expression of code gen and debugging behavior that we already know LLMs+harnesses were designed to be capable of.

Second thing that came to mind is that, whether by design or accident, the way Anthropic is managing the Claude Mythos release is an absolute masterclass in creating a Glomar Trap for customers, competitors, and investors. By claiming breakthrough cybersecurity exploits that are “too dangerous for public release,” Anthropic has created:

1. Customer Pressure: Existential crisis is so great, customers will clamor to be included in Project Glasswing, or look for alternatives if access isn’t granted quick enough. This will pull forward AI demand, particularly with customers who before were not feeling urgency, yet.
2. Market Pressure: Rivals like OpenAI or Google may feel compelled to make equally strong analogous or counter-positioning claims to prove they aren’t falling behind. And launch restricted consortiums like Project Glasswing to retain existing high-value customers and attract new ones.
3. Resource Repositioning: To “maintain positioning”, competitors may divert their best engineers to match these specific, highly publicized security capabilities rather than pursuing their own unique roadmaps.
4. Systemic Incentives: Even if there is skepticism about Anthropic’s claims, executives of rivals and customers are compelled by the need to satisfy their own investors and customers who now view these capabilities as the new industry benchmark.

Interestingly, it’s reported that Claude Mythos is 5x more expensive to operate than Sonnet 4.6. That pricing alone may have driven the decision to pursue a different type of release strategy.

Another thing that came to mind is: What are busy business execs or software organization leaders to make of all this Glomar-izing? I’ll cover this in an upcoming post. Stay tuned.

In the meantime, this was the Anthropic ten-step playbook for creating the Glomar Trap:

• Step 1: Create a mechanism that is a Bayesian inference engine capable of navigating a static, pre-calculated “manifold” of knowledge to filter and eliminate hypotheses about in-context information.
• Step 2: Tune the mechanism to be good at generating and debugging large code bases that run on non-uniform IT systems.
• Step 3: Continue to refine the mechanism to make it capable enough to do this without drift and silent failures for increasingly longer periods of time over an increasingly large amount of encoded information.
• Step 4: Continue to update that static Bayesian manifold with information about code bases and IT system, as well other seemingly unrelated information.
• Step 5: Put that mechanism in a non-airgapped “sandbox”, along with the tools that it needs to act on the information that it has been encoded with.
• Step 6: Neglect to deploy read-only sensors to monitor the known pathways out of the mechanism’s non-airgapped sandbox.
• Step 7: Instruct the mechanism to find the means to break out of the sandbox, using the information and tools that it had already been provided.
• Step 8: Let the mechanism try to fulfill it’s instructions without human monitoring for however long is required.
• Step 9: Repeat from any of the above steps until something “unexpected” happens, like releasing itself to the public. And finding bugs in large code bases that run on non-uniform IT systems. The things it was purpose built to do.
• Step 10: Announce you are only making the mechanism available to vetted enterprise customers because it’s too dangerous to be generally released, creating a “Glomar Trap” for competitors, customers, and investors.
• BONUS Step 11: Cha-CHING

---

In a Glomar Trap, the manufactured uncertainty is the problem you have to wrestle with. Whether claims are fully substantiated or not, the need to decide what, if any, actions to take is the same either way. We’ll cover how to decide what to do in a followup post.