Frontier AI Cyber Capabilities Double Every Four Months

Avatar photo

ByRyan Mitchell

May 4, 2026

New benchmarks from the UK AI Security Institute reveal rapid acceleration in offensive cyber capabilities as OpenAI and Anthropic models clear sophisticated network simulation tests.

The digital landscape has reached a critical inflection point as frontier artificial intelligence models demonstrate the ability to execute end-to-end offensive cyber operations. According to recent evaluations by the UK AI Security Institute (AISI), the velocity of progress in AI-driven cyber-offense is accelerating, with capabilities now doubling every four months. This represents a significant jump from the seven-month doubling rate recorded at the end of 2025.

Anthropic’s Claude Mythos Preview became the first model to clear the AISI’s 32-step “The Last Ones” range, a simulation designed to mimic corporate network takeovers. The model successfully completed the range in three out of ten attempts, maintaining a 73% success rate on expert-level tasks. OpenAI followed three weeks later with GPT-5.5, which achieved a near-identical profile by solving the simulation in two out of ten runs. While these tests were conducted without active human defenders, they indicate that AI is rapidly outstripping legacy, rules-based cybersecurity signatures.

The acceleration of these capabilities coincides with a massive influx of capital into the sector. OpenAI recently closed a $122 billion funding round at an $852 billion valuation, anchored by Microsoft, Nvidia, Amazon, and SoftBank. Not to be outdone, Anthropic secured an additional $40 billion from Google and $5 billion from Amazon. These investments highlight a strategic shift where frontier labs are evolving into massive infrastructure entities, backed by hundreds of billions in chip-supply and cloud-spend commitments.

On the geopolitical front, the traditional view of a “lag” in Chinese AI development is being challenged. Within a 12-day window, Chinese labs including Z.ai and MiniMax released open-weights coding models that achieved parity with Western frontier models on agentic engineering benchmarks. These models, such as GLM-5.1 and M2.7, are being offered at roughly one-third the inference cost of their American counterparts, suggesting that the competitive advantage in economically vital coding capabilities is narrowing.

In the robotics sector, Physical Intelligence released π0.7, a foundation model demonstrating zero-shot transfer to unseen tasks. The model successfully composed skills for complex kitchen workflows without specific retraining, mirroring the trajectory of large language models. This breakthrough suggests that robotics has transitioned into the foundation-model regime, potentially paving the way for more versatile autonomous systems in industrial and consumer settings.

Strategic realignments among major players are also reshaping the industry. Microsoft and OpenAI have renegotiated their partnership, moving toward a non-exclusive arrangement that allows OpenAI to diversify its compute sources while Microsoft ships rival models like Anthropic’s Opus 4.7. This shift signals the end of the exclusive platform-lab era as companies prioritize infrastructure diversification to maintain digital sovereignty in an increasingly volatile market.

Leave a Reply

Your email address will not be published. Required fields are marked *