Overview
A joint advisory from the Five Eyes intelligence alliance (US, UK, Canada, Australia, New Zealand) has warned of escalating cyber risks posed by AI models capable of autonomously compromising systems and networks. Security analyst Bruce Schneier frames this as the accelerating collapse of the historic link between skill and ability: for most of human history, causing sophisticated harm required deep expertise. Computers began decoupling the two, and AI is now widening that gap at an unprecedented pace.
The practical implication is stark: tasks that once required the expertise of groups like L0pht — who famously told the US Senate in 1998 they could take down the internet in 30 minutes — can increasingly be delegated to AI agents operating with minimal human direction.
Technical Analysis
Scheier draws a direct line from the “script kiddie” era — where pre-packaged exploit tools extended attack capability to unskilled users — to the current AI moment. Today’s AI systems, including non-frontier and open-source models, can:
- Autonomously enumerate and exploit network vulnerabilities
- Deploy ransomware payloads with minimal prompting
- Exfiltrate data and destroy systems
Critically, the article notes that smaller, locally-runnable open-source models are functionally equivalent to frontier models from OpenAI and Anthropic for many offensive tasks. These models are shared peer-to-peer, bypassing any centralised guardrails. Groups of models running in concert further amplify capability through agentic orchestration — each model handling a discrete phase of an attack chain.
The sociological argument is equally important: professional communities (medicine, engineering, security research) historically self-policed harmful knowledge through the socialisation that accompanies skill acquisition. AI removes that socialisation layer entirely, granting capability without the accompanying ethical framework.
Framework Mapping
- AML.T0047 (ML-Enabled Product or Service): AI is being weaponised as an attack platform, with models acting as autonomous offensive agents.
- AML.T0054 (LLM Jailbreak): Uncensored open-source models represent a structural jailbreak — guardrails are absent by design, not bypassed.
- AML.T0040 (ML Model Inference API Access): Local model deployment removes API-layer controls entirely.
- LLM08 (Excessive Agency): Agentic models executing multi-stage attack chains with minimal human oversight is a canonical excessive-agency risk.
- LLM05 (Supply Chain Vulnerabilities): The peer-to-peer distribution of uncensored models mirrors supply chain risks, with no provenance or safety guarantees.
Impact Assessment
The threat surface is effectively universal. Any organisation with internet-exposed infrastructure faces elevated risk from a dramatically enlarged pool of capable adversaries. The democratisation of offensive AI capability means incident volumes are likely to increase significantly, with attacks becoming faster, more varied, and harder to attribute. Defenders face an asymmetric burden: attack tooling improves continuously and is freely distributed, while defensive AI requires institutional investment.
Mitigation & Recommendations
- Invest in AI-assisted defence: The article explicitly states the only viable long-term response involves harnessing AI for defensive operations — threat detection, automated patching, and anomaly correlation at machine speed.
- Assume uncensored models are in adversary hands: Security controls should not assume attackers are constrained by commercial model policies.
- Monitor for low-skill, high-capability attack patterns: Autonomous AI attacks may lack the creative improvisation of skilled human operators — signature patterns may emerge.
- Engage with regulatory frameworks: The Five Eyes advisory signals incoming policy action; organisations should track and contribute to emerging AI security standards.