The Labs Are Grading Their Own AI Risk. Here's What SMBs Should Actually Do About It.

Scarlet Intelligence | Risk Intel Brief — July 2026

On July 1, Anthropic brought its newest model, Claude Fable 5, back online worldwide after a three-week suspension — and used the moment to publish something worth paying attention to well outside the AI industry: a detailed breakdown of how it grades the severity of AI "jailbreaks," built in partnership with Amazon, Microsoft, and Google.

It's a good-faith effort. It's also a preview of where AI risk governance is actually headed in 2026 — and it's not where most SMBs think.

What actually happened

Fable 5 and its more restricted sibling, Mythos 5, launched June 9. Three days later, new U.S. export controls forced Anthropic to suspend both models for every user globally — not because of a flaw, but because real-time nationality verification wasn't possible at the time, and Anthropic chose to pull access rather than risk non-compliant use. Access was restored July 1, once the Commerce Department's controls were lifted.

On redeployment, Anthropic published two things:

• A four-tier classifier system that sorts cybersecurity-related requests by risk — from Prohibited (ransomware, wipers, industrial-control-system sabotage, malware development, always blocked) down to Benign Use (secure coding, patch management, SOC analysis, incident response — permitted, occasionally over-flagged).

• A proposed severity framework — built with Amazon, Microsoft, and Google as part of an effort called Glasswing — that scores jailbreak techniques on a five-band scale from CJS-0 (Informational) to CJS-4 (Critical), using four factors: capability gain, breadth, ease of weaponization, and discoverability. Industry press is already calling it the CVSS equivalent for AI — a shared severity language the way the security world already has one for software vulnerabilities.

That's a real step forward. It's also, structurally, the labs writing their own report card.

Why "the labs grading themselves" isn't a criticism — it's the current state of the world

There's no federal AI jailbreak standard. There's no third-party auditor scoring these frameworks the way a C3PAO scores CMMC controls. What exists right now is a handful of frontier labs, working together voluntarily, publishing their own severity criteria and their own classifier design decisions — genuinely useful, genuinely well-intentioned, and entirely self-reported.

That's not unique to Anthropic. It's the pattern across the entire AI risk landscape in mid-2026: labs are moving faster than regulators, so the labs are setting the vocabulary regulators will eventually borrow. Worth noting — and this cuts against any single-vendor picture — Amazon's own testing during this same disclosure found the underlying jailbreak behaviors Anthropic reported exist across every frontier model they tested, not just Fable 5. This is a class-level problem across the AI industry, not a one-vendor gap you can shop your way around.

What this actually means if your business touches AI

Whether you're using an LLM API in your product, or just have employees using AI coding and support tools daily, three things from this disclosure are directly relevant — regardless of which vendor you use:

1. A vendor's safety framework is a data point, not a certification.

The same principle that applies to NIST frameworks and SOC 2 applies here: a lab publishing its own severity criteria tells you how it thinks about risk. It doesn't verify that the controls work as described, and it isn't audited by anyone outside the lab. Treat it the way you'd treat a vendor's self-attested security questionnaire — useful context, not proof.

2. Safety tuning has operational side effects, and you should expect them.

Anthropic said plainly that the retrained classifier flags more legitimate coding and debugging requests than before — a deliberate trade-off, with blocked requests quietly falling back to a less capable model. If your team leans on AI tools for security work, code review, or incident response, a vendor's safety update can change your workflow overnight without warning. That's a vendor-risk event, not just a product update.

3. AI vendor dependency carries geopolitical risk, not just technical risk.

Fable 5 and Mythos 5 were suspended for every user on the planet within three days of launch because of an export control, with no advance notice to end users. If any part of your operation depends on a frontier AI vendor, that's now a documented, recent example of access disappearing overnight for reasons that have nothing to do with your business or the vendor's product quality.

The practical takeaway

The four-factor language in Anthropic's proposed framework — capability gain, breadth, ease of weaponization, discoverability — is genuinely useful, and SMBs don't need to wait for an industry standard to start using it internally. Ask it about your own AI tool stack: if an employee misused this tool, what capability would they gain, how broadly could it be replicated, how easy would it be to weaponize, and how discoverable is the risk today? That's a five-minute exercise that tells you more about your actual exposure than checking whether your vendor published a nice-looking framework.

The labs are doing something worth encouraging. They're also, unavoidably, marking their own homework. The businesses that treat lab-published safety frameworks as one input into an ongoing risk picture — instead of a checkbox that says "handled" — are the ones who won't be caught flat-footed the next time a vendor's classifier changes, or a model disappears from their stack without warning.

---

Scarlet Risk helps SMBs maintain continuous risk intelligence across their AI and vendor stack — not a one-time certification, an ongoing read on where the real exposure sits.