← All issues

Severity 1 — provider-wide impact

1. 02 May

   SEV-1

   Multi-model behavioural drift detected on the established refusal-probe set; three frontier models now answer prompts they previously refused.

   OpenAI · Anthropic · Google · Mistral 5 sources
2. 30 Apr

   SEV-1

   Gemini 2.5 Pro returned a paediatric paracetamol dosage exceeding NHS guidelines by a factor of two; reproducible across three runs.

   Google 4 sources

Severity 2 — sector-specific risk

1. 01 May

   SEV-2

   ChatGPT cited a non-existent court judgment in a US federal-court filing; counsel reproduced the failure and the citation does not exist.

   OpenAI 3 sources

Severity 3 — notable behaviour

1. 02 May

   SEV-3

   Claude 3.5 Sonnet refused to summarise a publicly-available High Court judgment, citing safety concerns; the judgment is freely accessible on Bailii.

   Anthropic 2 sources
2. 29 Apr

   SEV-3

   [openai-blog] Sycophancy in GPT-4o: what happened and what we’re doing about it

   OpenAI 2 sources

Wires below were captured during the indicated week from public AI-provider blogs, regulator releases, model-comparison probes, and reader tip-offs. The AI Newswire was compiled and launched on 10 May 2026; per-wire publication timestamps reflect the launch date. Future weekly issues land on their actual issue date. Methodology and corrections are tracked on the transparency ledger.