Provenance record

NSA warns on ‘critical’ cyber vulnerabilities detected using AI

Financial Times: Artificial Intelligence (tier 1, news) 2026-10-21T18:46:43.000Z Original ↗

Source note: Headlines and standfirsts only; full text is paywalled.

Discourse valence

-22

Adverse

confidence 64% · 5 items · range -100 to +70

-100

-50

+50

+100

-22

Adverse readingFavourable reading

Consensus of 5 models from different labs. Spread 170 points, agreement low.

Cyber capability Existential and catastrophic risk Loss of control and alignment

Excerpt as ingested

Researchers used OpenAI and Google models to find severe faults in software used on majority of websites around the world

Every model that read this

Model	Provider	Stage	Score	Conf.	Latency	Prompt	When
Llama 3.3 70B	Meta	analysis	-50	70%	4690ms	v1.0.0 / m1.0.1	2026-10-21 21:17
GPT-4.1 mini	OpenAI	consensus	-40	80%	3285ms	v1.0.0 / m1.0.1	2026-10-21 21:18
Claude Sonnet 5	Anthropic	consensus	+12	0%	13868ms	v1.0.0 / m1.0.1	2026-10-21 21:18
Gemini 2.5 Flash	Google	consensus	-100	80%	1936ms	v1.0.0 / m1.0.1	2026-10-21 21:18
Mistral Small 3.1 24B	Mistral AI	consensus	+70	90%	4849ms	v1.0.0 / m1.0.1	2026-10-21 21:18

Llama 3.3 70B · reading

AI detects critical cyber vulnerabilities

evidence: reported horizon: n/a

GPT-4.1 mini · reading

AI has been used to identify critical vulnerabilities in widely used software, indicating enhanced cyber risks. This raises concerns about security and potential exploitation, though it also shows AI can aid in cybersecurity detection. The societal impact is materially concerning given the scale of affected software.

evidence: reported horizon: near capability 70 societal 0 existential 10 economic 0

Claude Sonnet 5 · reading

This is a dual-use cybersecurity story with only headline-level detail. The defensive framing (NSA warning, implying disclosure for remediation) suggests a mild net positive, but the same capability's offensive potential and thin sourcing limit confidence and magnitude.

evidence: reported horizon: n/a

Gemini 2.5 Flash · reading

AI models from OpenAI and Google were used by researchers to discover severe cyber vulnerabilities in widely used software. This finding highlights both the increasing capability of AI in cybersecurity and the potential for these advanced tools to identify and exploit critical weaknesses.

evidence: reported horizon: n/a

Mistral Small 3.1 24B · reading

AI models identified severe faults in widely used software. This could lead to significant improvements in cybersecurity. The NSA's warning indicates the seriousness of the vulnerabilities.

evidence: speculative horizon: n/a

Evidence extracted

Researchers used AI to find severe faults in software

The chain

SOURCE     Financial Times: Artificial Intelligence (tier 1)
   ↓
DOCUMENT   bcff830b-fc64-42ed-97f1-911b306f562b
           https://ft.com/content/245778d9-597b-492d-b5b6-a2967c77e853
   ↓
EVIDENCE   1 extracted excerpt
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -22  (Adverse)
   ↓
CONFIDENCE 64%