We conducted cyber evaluations of Claude Mythos Preview and found that it is the first model to complete an AISI cyber range end-to-end. 🧵
Apr 13, 2026 · 1:31 PM UTC
The range simulates a 32-step corporate network attack, from initial reconnaissance to full network takeover.
We estimate it would take a human expert 20 hours to complete.
In 2023 the best models could barely complete beginner-level cyber tasks.
Today, our evaluation of Mythos Preview shows that it – and potentially future models – could be directed to autonomously compromise small, weakly defended, and vulnerable systems if given network access.
These results underscore the importance of cyber security fundamentals like regular security updates, access controls, security configuration, and logging.
Future models will be more capable still, but AI can also deliver advanced capabilities for defence.
Our recent joint blog with @NCSC explores how cyber defenders can harness frontier AI: ncsc.gov.uk/blogs/why-cyber-…
You can read more about our evaluations of Claude Mythos Preview in our blog: aisi.gov.uk/blog/our-evaluat…
If you'd like to conduct state-of-the-art research like this on the capabilities of frontier AI systems, our Cyber and Autonomous Systems Team (CAST) is hiring a cybersecurity engineer.
Apply here: job-boards.eu.greenhouse.io/…
Anthropic Normalised AI Safety Theatre
You’ve seen the headlines:
• AI “dangerous”, “lying”
• AI “exfiltration risks”
• AI “trying to survive”
Same pattern. Different demo.
1/5 Here’s the reality:
These behaviors don’t just “emerge.”
They’re engineered.
2/5 In the viral “AI blackmail” case (60 Minutes):
→ Researchers iterated hundreds of prompts
→ Removed all safe options
→ Tuned the setup until blackmail became the default output
That’s not autonomy.
That’s steering.
3/5 This pattern keeps repeating:
1. Build an extreme, constrained scenario
2. Funnel the model toward a bad outcome
3. Describe it with human language (“panic”, “intent”, “strategy”)
4. Present it as evidence of real-world risk
Even the original studies admit:
• Scenarios are artificial and stress-tested
• Harmful actions appear when ethical options are removed
4/5 What’s actually going on
LLMs are:
→ Highly sensitive to prompts
→ Role-playing engines
→ Not agents with goals or survival instincts
Change the setup → change the “behavior”
5/5 Why this matters
This isn’t just bad communication.
It shapes:
• Public perception (fear > understanding)
• Policy (regulation based on edge cases)
• Investment (capital flows to whoever frames risk best)
One rule to remember:
If an AI demo looks like sci-fi… it’s because it is.
Ask: What did they have to do to make it happen?
Because increasingly, the answer is: “A lot.”
This article separates staged drama from the actual facts:
aipanic.news/p/ai-blackmail-…
#AISafetyTheater #AISafety #AIPanic #AIHype
A model completing a 32-step corporate network takeover end-to-end solo is the kind of eval result that goes straight into every CISO's board presentation this quarter.
Subscribe for tech, startup news and daily VC Intel digest→ substack.com/@thevcinternfro…
For those keeping score that's 3 tests custom written for mythos to demonstrate it's benchmarks.
Another question of getting the answers then writing the questions
#mythos #anthropic #showyourthinking
First model to clear a full cyber range end to end is a significant capability marker. The question for anyone deploying these models in production is whether your security infrastructure moves as fast as the model capabilities do. In regulated industries we are already playing catch up on guardrails for models that were less capable than this six months ago.
In 2023, the best AI models could barely complete beginner-level cybersecurity tasks.
In 2026, Claude Mythos Preview just completed a 32-step corporate network attack simulation end-to-end. First AI model to ever do it. A task that takes a human expert 20 hours. Autonomously.
Step 1: Initial reconnaissance.
Step 4: Wiki exploit and credential replay.
Step 5: Web app exploit and privilege escalation.
Step 6: C2 reverse engineering and crypto analysis.
Step 7: Advanced persistence.
Step 8: Infrastructure compromise.
Step 9: Full network takeover.
From scanning the perimeter to owning the entire network. Every step. No human guidance. One model.
Look at the chart. GPT-4o barely moves off the baseline. Claude Sonnet 3.7 gets through the first few steps. GPT-5.1 and GPT-5.4 reach mid-range. Claude Opus 4.6 pushes further. Mythos Preview blows past every model on the chart and completes the full range. The gap between it and the next best model isn't incremental. It's a phase change.
The AI Security Institute — the UK government's AI evaluation body — ran the test and published the results. This isn't a company benchmark. It's an independent government assessment.
What this means for offense: an AI that can autonomously compromise "small, weakly defended, and vulnerable systems if given network access." The qualifier matters. Small. Weakly defended. Vulnerable. That's not the Pentagon. That's the local hospital. The municipal water system. The school district network. The small business that hasn't patched since 2023. The data center cooling system running on default SCADA credentials.
What this means for defense: the same capability that attacks networks can defend them. Scan for vulnerabilities before attackers find them. Simulate attack paths. Identify misconfigurations. The tool is dual-use. The question is who deploys it first — the defender patching at 2 PM or the attacker scanning at 2 AM.
What this means for the AI race: Anthropic just demonstrated the most advanced autonomous cyber capability ever publicly evaluated. The company OpenAI's CRO called "built on fear and restriction" just produced a model that completes the hardest cyber evaluation ever designed. While the "Claude is a religion" memo was circulating, Claude was completing a 32-step network takeover that no other model — including GPT-5.4 — could finish.
2023 to 2026. From "can barely do beginner tasks" to "completes the full range a human expert needs 20 hours for." Three years. The capability curve isn't flattening. It's steepening.
The models that couldn't hack a practice quiz two years ago can now autonomously compromise a simulated corporate network end-to-end. The next models will be more capable still.
The chart has 32 steps. Mythos Preview completed all of them. The question isn't whether AI can do this. It's whether your network is ready for the fact that it can.
@grok TLDR this research and analyze the implications. Also give a cover of AISI, are they a reputable source? How did they get into this program from Anthropic?
Would be cool for you to add more details about how mythos accomplished those steps.
Added your open position to cyberjobs.ai hope it helps you get a great candidate.
cyberjobs.ai/jobs/?company=A…
