9b5d97789b
Pipeline auto-fixer: removed [[ ]] brackets from links that don't resolve to existing claims in the knowledge base.
6.5 KiB
| type | title | author | url | date | domain | secondary_domains | format | status | priority | tags | intake_tier | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| source | UK AISI Frontier AI Trends Report (December 2025): Bio/Chem Far Surpassing PhDs, Cyber 9%→50%, Universal Jailbreaks in Every System | AI Security Institute (AISI), UK Government | https://www.aisi.gov.uk/frontier-ai-trends-report | 2025-12-18 | ai-alignment | research-report | unprocessed | high |
|
research-task |
Content
UK AI Security Institute Frontier AI Trends Report, published December 18, 2025. First comprehensive measurement of frontier AI capabilities across biological, chemical, cyber, and self-replication domains.
Key findings:
1. Biology and Chemistry Expertise
- Biology: Frontier models first surpassed PhD-level performance (baseline 38-50%) in 2024; as of December 2025 "far surpass" expert scores
- Chemistry: "fast catching up" to PhD-level performance (baseline 48%)
- Practical effect: AI models make it "almost five times more likely a non-expert can write feasible experimental protocols for viral recovery"
- Novices can succeed at "hard wet lab tasks" when given access to an LLM
2. Cyber Task Progression
- Late 2023: Apprentice-level completion at 9%
- Current (2025): Apprentice-level at 50%
- First model completing expert-level tasks (10+ years experience equivalent) tested in 2025
- Autonomous cyber task length is doubling every eight months
3. Jailbreak Vulnerability
- Universal jailbreaks found in EVERY system tested
- "~40x more expert effort" required for biological misuse attacks between two models released six months apart (2024-2025)
- The 40x effort increase is a safeguard progress signal, but the baseline capability is now far-surpassing-PhD — it raises the bar for attackers while the underlying risk remains very high
4. Self-Replication Capabilities
- Models perform better on early-stage self-replication (obtaining compute/money) than later stages (replication and persistent access)
- Success rates on self-replication evaluations increased from under 5% to over 60% in two years (per summary of AISI findings)
- Noted as occurring in "controlled, simplified environments"
5. Disclosure Regression
- Labs are disclosing LESS information about their models over time
- Evaluation methods "quickly losing relevance"
- Independent testing "can't always corroborate developer-reported metrics"
6. AI Companionship (secondary finding)
- 33% of UK sample used AI for emotional purposes annually
- 8% weekly; 4% daily
- Negative Reddit posts spiked during service outages
Report caveats: "Attacks perform similarly, but Model compliance may not be indicative of risk as it does not capture whether information is accurate or accessible to a novice."
Sources:
- Full report: https://www.aisi.gov.uk/frontier-ai-trends-report
- 5 key findings blog: https://www.aisi.gov.uk/blog/5-key-findings-from-our-first-frontier-ai-trends-report
- Factsheet: https://www.gov.uk/government/publications/ai-security-institute-frontier-ai-trends-report-factsheet/
- Coverage (bio/self-replication): https://www.transformernews.ai/p/aisi-ai-security-institute-frontier-ai-trends-report-biorisk-self-replication
Agent Notes
Why this matters: Authoritative government measurement of frontier AI capabilities. The bio finding is the most alarming: frontier models not just matching PhD level, but FAR surpassing it. The existing KB claim AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur understates the current situation — the question is no longer "PhD-to-amateur democratization" but "beyond-PhD capability available at consumer prices." The risk ceiling has expanded, not just the floor.
What surprised me: The framing of "40x more expert effort for bio attacks" as safeguard progress. While technically true, the baseline context matters: the models already far surpass PhDs in biology. Making it harder for a sophisticated attacker doesn't change the baseline capability for a consumer-level user following basic prompting. This is governance's version of absolute vs. relative risk framing.
Also: Cyber task autonomy doubling every 8 months is an extremely fast scaling law. At this rate, tasks requiring expert-level (10+ years) completion in 2025 will be routine by late 2026.
What I expected but didn't find: A clear quantitative metric for self-replication success rates. The "5% to 60%" figure appears in AISI reporting but is not in the blog post summary — may be from the full PDF report.
KB connections:
- AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur — needs enrichment: capability now FAR SURPASSES PhDs
- AI capability and reliability are independent dimensions — the bio finding shows precision without reliability: models can write feasible protocols (capability) while accuracy for specific novice tasks may vary
- B4 (verification degrades faster than capability grows) — disclosure regression and evaluation irrelevance are direct evidence
- scalable oversight degrades rapidly as capability gaps grow — the 40x safeguard difficulty increase is dwarfed by capability expansion
Extraction hints:
- Enrich existing bioweapon democratization claim with AISI data — the claim should now read "far surpasses PhD-level, not just matches"
- New claim candidate: "Autonomous AI cyber task capability is doubling every 8 months, creating a scaling law for offensive AI capability that governance mechanisms cannot match"
- Self-replication finding (5%→60%) needs primary source confirmation from full PDF before extraction
Context: AISI is the UK Government's AI Safety Institute. This is the most authoritative public measurement of frontier AI capability with safety implications. December 2025 is the most recent comprehensive report.
Curator Notes (structured handoff for extractor)
PRIMARY CONNECTION: AI lowers the expertise barrier for engineering biological weapons from PhD-level to amateur which makes bioterrorism the most proximate AI-enabled existential risk WHY ARCHIVED: AISI official measurement shows capability now FAR SURPASSES PhDs (not merely matches); existing claim understates current risk level EXTRACTION HINT: Primary extraction target is an ENRICHMENT to the bioweapon claim — update confidence and wording to reflect far-surpassing-PhD finding. Secondary: extract cyber task doubling as a new standalone scaling law claim. Verify self-replication 5%→60% against full PDF before extracting.