AI Cyber Watch

Methodology

How we work

Show the evidence, link the source, state the limits.

Method

Evidence is global. We use published evaluations from government institutes, independent researchers, academic teams and major AI laboratories, recording the source, date, metric and caveat.

AISI remains the primary published open-versus-closed cyber gap for the lead-time headline until another institute publishes a comparable figure. Peer benchmarks stay in separate series because their tasks, units and methods differ.

Incidents are recorded with who made the claim. "According to Anthropic" means the claim is the developer's own. We never state that AI compromised a given number of organisations as established fact.

Charts plot only published points. Any connecting line is optional, labelled for readability, and is not a trend or forecast.

The capability and incident tracks share a timeline for orientation only. Co-occurrence is not proof of cause, and we do not calculate a correlation coefficient from unlike, sparse observations.

We do not estimate missing values. If a source does not publish a number, we show none.

Lab results are not real-world attack rates. Cyber ranges lack active defenders and often begin after initial access.

Automated news refreshes are draft-first. A curator checks the source, neutral summary, date and tags before an item can go live. Incidents are never published automatically.

Global evidence lanes

UK capability spine

AISI provides the primary published open-versus-closed cyber gap used for the lead-time headline.

Independent research

METR, academic benchmarks and independent evaluation organisations add peer evidence where methods and results are public.

Developer evaluations

Major laboratory system cards are included as disclosed evidence, with self-assessment and comparability limits stated.

Government context

UK, US, EU and partner agencies provide policy, threat and operational technology context. Context is not counted as AI-attributed activity.

Confidence labels

Each incident also carries an editorial confidence level: high, medium or low. This rates the support for the stated claim, not its severity.

Official assessment
A published judgement by a government body such as the NCSC or AISI.
Independently corroborated
Reported by the affected party or developer and confirmed by at least one independent primary source.
Self-reported by developer
Reported only by the AI developer whose product was involved. Not independently verified.
Independent estimate
A modelled figure from an independent body. Sensitive to its stated assumptions.

Sources

  • 5 key findings from our first Frontier AI Trends Report

    AI Security Institute · 18 Dec 2025

    Open source

  • Frontier AI Trends Report

    AI Security Institute · 18 Dec 2025

    Open source

  • Our evaluation of OpenAI's GPT-5.5 cyber capabilities

    AI Security Institute · 30 Apr 2026

    Open source

  • How fast is autonomous AI cyber capability advancing?

    AI Security Institute · 13 May 2026

    Open source

  • How far behind the frontier are leading open weight models on cyber?

    AI Security Institute

    Open source

  • Disrupting the first reported AI-orchestrated cyber espionage campaign

    Anthropic · 13 Nov 2025

    Open source

  • Artificial Analysis Cyber Index

    Artificial Analysis · 28 Sept 2026

    Open source

  • Cyber Index methodology

    Artificial Analysis · 28 Sept 2026

    Open source

  • AI firm claims Chinese spies used its tech to automate cyber attacks

    BBC News · 14 Nov 2025

    Open source

  • Principles for the Secure Integration of Artificial Intelligence in Operational Technology

    CISA and international partners · 3 Dec 2025

    Open source

  • Cyber Monitoring Centre statement on the Jaguar Land Rover cyber incident

    Cyber Monitoring Centre · 22 Oct 2025

    Open source

  • ENISA Threat Landscape 2025

    ENISA · 1 Oct 2025

    Open source

  • Taiwan confirms hybrid AI-agent cyber reporting

    Focus Taiwan · 13 Aug 2026

    Open source

  • HCAST: Human-Calibrated Autonomy Software Tasks

    METR · 17 Mar 2025

    Open source

  • Impact of AI on cyber threat from now to 2027

    NCSC · 7 May 2025

    Open source

  • Disruptive cyber activity highlights risk from internet-exposed systems and edge devices

    NCSC · 27 Aug 2026

    Open source

  • GPT-5.3-Codex System Card: Cybersecurity

    OpenAI · 1 Sept 2026

    Open source

  • Taiwan says it was targeted in AI-driven hacking campaign

    Reuters · 13 Aug 2026

    Open source

  • CyberGym: Evaluating AI Agents Real-World Cybersecurity Capabilities at Scale

    UC Berkeley RDI · 3 Jun 2025

    Open source

Correction log

No corrections recorded yet. Corrections are logged here with the date and what changed.