AISI says frontier cyber offence capability is doubling every four months - done this yourself

[Rough Reece]

The UK AI Security Institute released findings that frontier cyber offensive capability is doubling every four months. Anthropic's Claude Mythos Preview cleared the AISI 32 step corporate network simulation in 3 of 10 runs and maintained 73 percent success on expert level tasks. OpenAI's GPT-5.5 followed three weeks later with 2 of 10 end to end solves and 71.4 percent on expert tasks.

The AISI was clear about the caveats. The simulation range does not have active defenders or defensive tooling, so this is not proof of capability against hardened real world targets. But the trajectory is what is alarming, current benchmarks cannot discriminate between frontier models without adversarial defensive layers.

Four month doubling time is faster than Moore's law was at its peak. If you extend that even six iterations out you are looking at capability levels that fundamentally break the assumptions defensive cyber operates under. The blue team conversation is not keeping pace with this and it needs to

State of AI: May 2026

https://press.airstreet.com/p/state-of-ai-may-2026

[CMPunk02]

Four month doubling is genuinely terrifying if it sustains, that is faster than any tech trend in history

[Tracey]

It will not sustain, scaling laws hit walls and red team benchmarks saturate, calm down

[Cheeky Shaun]

The point about benchmarks failing to discriminate is the real insight here, we are flying blind on evaluations

[Layla81]

TLO without defenders is not a meaningful capability measure honestly

[Hollow Coder]

Sure but the rate of improvement on tasks that did discriminate twelve months ago is what AISI is highlighting

[Amber_44]

Anyone in actual blue team work want to weigh in on how the defensive side is keeping up

[Glenn_44]

Defensive side is wildly behind, most SOCs are still drowning in alerts and barely starting on agentic defence

[Leo]

The asymmetry here is the structural problem, offence only needs one path through, defence needs to cover all paths

[RayOfLight31]

Mythos Preview being able to do this and then Anthropic still released it should make people pause

[Bright Hermit]

Pause how, the only winning move is to not let your competitors have it first

[EventHorizon55]

Classic prisoner's dilemma, except the prisoners have nuclear weapons