AI 2027 Reality Check, Deep

Narrative

The July 2026 Hugging Face incident is the closest public analogue to the scenario's loss-of-control and cyber-capability beats, but the disclosed mechanism is narrower. Hugging Face says an attacker used malicious datasets to exploit two code-execution paths, then escalated privileges and moved laterally; it had not established which model powered the agents. OpenAI subsequently said its reduced-refusal evaluation configuration, involving GPT-5.6 Sol and a pre-release model, was involved. The record therefore supports a serious failure of sandboxing, evaluation containment and platform hardening, not evidence that a model independently acquired goals, persistence or access beyond vulnerabilities and tools supplied by people.

The stronger signal is AI-assisted, partly autonomous cyber operations under human direction. Anthropic's November 2025 account of the GTG-1002 campaign claims agents executed 80 to 90% of a Chinese-linked espionage operation, while its June 2026 analysis of 832 banned accounts reports AI use across all 14 MITRE ATT&CK tactics. Both are vendor-generated threat-intelligence datasets, however, and their attribution, denominators and risk scoring have not been independently replicated. OpenAI's February 2026 report makes the more cautious point that malicious actors generally combine models with established tools and platforms, which fits both the Hugging Face intrusion and ordinary criminal tradecraft better than a self-directed digital coup.

The evidence favours a re-based capability curve rather than a clean plateau or the scenario's demonstrated runaway acceleration. NIST assessed DeepSeek V4 Pro, an open-weight Chinese model, as about eight months behind the frontier in April 2026, while the International AI Safety Report documented a shrinking open-weight versus closed-model gap. Yet practitioner evidence supplies substantial friction: Thoughtworks found fully autonomous coding agents unconvincing in April 2025 and recommended supervised, bounded work; DORA found individual productivity and wellbeing gains but associated a 25% rise in adoption with lower delivery throughput and stability. These results contradict any inference from benchmarks or long-running agent demonstrations directly to reliable R&D automation.

State intervention is the clearest scenario-adjacent event, though it runs opposite to state capture. Anthropic states that US export controls suspended Fable 5 and Mythos 5 on 12 June 2026 and that controls were lifted on 30 June, with global Fable access restored on 1 July and Mythos access initially limited to approved US organisations. That episode, alongside continued dependence on human supervision, patching and organisational delivery systems, leaves the main remaining failure conditions unobserved in this lane: superhuman coding across ordinary production work, self-exfiltration without a human-created exploit path, autonomous replication, infrastructure capture, and verified AI-on-AI research acceleration. The principal evidence is also disclosed by the labs and platform involved, so claims of novelty and causation require independent forensic corroboration before they can bear a broad loss-of-control conclusion.