Fleet has joined Anthropic’s Project Glasswing to test Claude Mythos Preview for defensive security work. The collaboration will put a frontier AI model against practical device-management tasks, where a rushed fix can cause its own trouble.
Anthropic Watch analysis
What happened
ChannelLife UK reported on 29 September that Fleet has early access to Claude Mythos Preview through Project Glasswing, Anthropic’s controlled-access programme. Fleet plans to examine how the model could help secure device-management systems, mobile-management protocols and the infrastructure behind them.
The work is focused on operational tasks: assessing endpoint state, prioritising remediation and helping teams respond to security risks more quickly. Fleet says it intends to share lessons and practical guidance from the research publicly.
Fleet also cited its own research, which found that 79% of organisations take more than a day to deploy critical patches for browsers, operating systems and extensions. That figure is Fleet’s, and the article does not provide further detail about the research methodology.
Why it matters
Security teams face a blunt timing problem: a vulnerability may need urgent attention, but applying changes across a mixed fleet of laptops, phones and other devices takes care. AI that helps teams spot the right issue or set priorities could be useful. AI that acts too freely could turn a quick fix into a fresh incident.
The project’s value will depend less on the phrase “frontier AI” than on what it can safely do in real workflows, and whether Fleet’s findings are useful beyond its own products.
Our read
This is a concrete test of AI in defensive security, not proof that AI has made patching faster or safer. The sensible promise is assistance with assessment and prioritisation, with humans still accountable for changes that affect devices at scale. Fleet’s plan to publish its lessons gives the wider security community something practical to judge, if those lessons arrive with enough detail to be useful.
What to watch
- Whether Fleet publishes results, including limits and failures as well as successes.
- Which security tasks the model can support reliably, and which remain under human control.
- Whether the findings apply beyond Fleet’s own device-management environment.
Discussion spark: Would you trust an AI system to prioritise or apply security fixes across a company’s devices, and where should human approval be mandatory?
Sources and evidence
- Fleet joins Anthropic project to test AI for security – ChannelLife UK (29 September 2026, 08:15 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.