Key Points
- Anthropic released Claude Fable 5, a public-facing version of its restricted Mythos model, adding safeguards to block high-risk uses.
- The system routes sensitive queries (cybersecurity, biology) to a weaker model (Opus 4.8) and is priced roughly two times higher, reflecting additional compute.
- IANS Faculty recommend treating AI-driven vulnerability discovery as already operational on both sides -- tighten asset visibility and compress patch SLAs to hours or days.
Anthropic Releases ‘Fable 5’ as Guardrailed Version of Mythos
On Tuesday, Anthropic released Claude Fable 5, the first widely available version of its Mythos-class AI model, after initially withholding the original system due to concerns it could be used to discover and exploit software vulnerabilities at scale.
Fable 5 runs on the same underlying architecture as Mythos but adds hard safety limits. In high-risk areas such as cybersecurity (e.g., exploiting software bugs) and biology, the model blocks responses or routes queries to Anthropic’s earlier Claude Opus 4.8 system.
Anthropic said it added these controls to enable a broad release without exposing the full offensive potential of the model. The company previously limited Mythos access to a small group of infrastructure and cybersecurity organizations under its Project Glasswing program.
Anthropic said the model outperforms prior systems across software engineering and knowledge work and represents a “significant jump” in capability over Opus 4.8. Its pricing is roughly double Opus 4.8, reflecting the higher compute.
Big Picture
Anthropic is moving forward with releasing a Mythos-level model, but doing so by layering controls on top of the capability rather than changing the capability itself.
"There is a nice positive takeaway here in that the community has realized that AI has to be governed...like privileged access in some ways -- tiered, monitored, justified, and revoked (or maybe redirected in this case) when misused.” Summer Fowler, IANS Faculty.
That approach depends on guardrails holding up in practice. The controls can redirect or block certain requests, but they sit at the interface level and rely on the assumption that misuse can be contained.
"Routing the sensitive asks back to Opus 4.8 raises the cost and the noise of trying to misuse it, and that's worth something. But a guardrail that lives at the prompt is a speed bump. The capability still exists in the model class, and a determined actor breaks the task into pieces that never trip the filter.” George Gerchow, IANS Faculty.
That leaves open the likelihood that Mythos’ underlying capability will still be used more broadly, particularly in identifying vulnerabilities across software at scale.
"One of Mythos' most significant advantages is its ability to identify vulnerabilities in software without requiring access to source code. There are effectively no guardrails preventing Mythos Lite from evolving into the most powerful decompiler the industry has ever seen. Even in its current form, it represents a step-change improvement in efficiency for vulnerability discovery.” Aaron Turner, IANS Faculty.
IANS Faculty Recommendations
- Tighten asset and dependency visibility: Make sure you have a clear, current view of systems and software so newly discovered vulnerabilities can be quickly understood and acted on.
- Compress the find-to-fix cycle: Reduce the time between discovery and remediation so issues are addressed before they can be exploited.
- Plan for guardrail bypass: Expect actors to break tasks into smaller steps or use multiple models to work around controls.
- Stay disciplined on resourcing: Avoid shifting budget toward vulnerability discovery unless you are confident it will not impact recovery or other critical capabilities.
Authors & Contributors
Dan Maloof, Editor-in-Chief, IANS News
Summer Fowler, IANS Faculty
George Gerchow, IANS Faculty
Aaron Turner, IANS Faculty
Although reasonable efforts will be made to ensure the completeness and accuracy of the information contained in our News & blog posts, no liability can be accepted by IANS or our Faculty members for the results of any actions taken by individuals or firms in connection with such information, opinions, or advice.