All editions
AI News5 items

AI News, Monday 26 October 2026

1. Stolen GitHub tokens are dumping repository secrets.

Know · Wiz Research has detailed Gheist, an automated campaign it has tracked since 15 September. With a stolen GitHub token, the tool pushes a branch named ci-sync- plus 6 characters and a workflow file called sync-check.yml to every repository the token can reach. The workflow prints your secrets in an encoded form that slips past GitHub’s log masking, downloads the log, then deletes the run seconds later. As of 24 October, Wiz counts more than 13,500 planted branches across more than 850 accounts. GitHub holds these runs for approval on public repositories, but not on private ones; in some companies, the stolen cloud keys were used hours later. Wiz links the tokens it could trace to infostealer logs and older supply-chain leaks. What to do: search your repositories for ci-sync- branches today, and replace long-lived tokens with fine-grained ones that expire. Source

2. Model copies working as a team beat Claude Fable 5 at coding.

Try · A public GitHub project called GVS5H, shared on Hacker News on 25 October, tests a simple setup: several fresh copies of one model split a coding problem and coordinate through shared files (a plan, notes, the current solution). On 100 recent hard problems from LiveCodeBench, an orchestrated Qwen3.8 Flash Next solved 93.0%, against 90.4% for Claude Fable 5, at 9% of the cost. A Qwen3.8-27B model served locally went from 66.8% to 92.4%. The authors warn that gains are not universal, and some models did not improve. Why it matters: a shared plan file and fresh context per subtask is a pattern you can copy in your own agent setups. Clone the repo, add an API key, and compare the single-call baseline with the manager mode. Source

3. A hospital worker uploaded 3,182 patients to a public AI tool.

Know · Johns Hopkins Bayview Medical Center, in Baltimore, started mailing letters on 22 October to 3,182 patients. A staff member had uploaded a spreadsheet of patients treated from May to September 2026 to a public AI tool the hospital had not approved. An internal review found it on 21 September. The file held names, record numbers, treatment dates, medications and provider names, but no Social Security or bank details. The hospital asked the tool’s maker to delete the data and is retraining staff on approved tools. Why it matters: it is a fresh, concrete case of shadow AI for your governance workshops, and it shows clients why staff need an approved tool and clear rules on what data goes in. Source · Source 2

4. Why OpenAI and Anthropic cut prices by up to 80%.

Know · Data scientist Max Woolf published on 25 October a timeline of model price cuts since July. OpenAI cut GPT-5.6 Luna by about 80% on 30 July. On 22 September, Anthropic priced Opus 5.5 at $4 input and $20 output per million tokens, down from $5 and $25 for Opus 5. Woolf notes that in agentic coding, reading cached context makes up 96 to 99% of token spend, so cuts on cache prices count the most. His main explanation is competition: switching models is easy, and one open-weight DeepSeek model now runs on 29 hosting providers, up from 13 in May. Why it matters: the running cost of the automations you build for clients keeps falling, so check current prices before each new quote. Source

5. A 10-year-old image model beats frontier LLMs on skin cancer.

Know · Magnus Lynch, a dermatologist at Guy’s Hospital in London, published a test on 25 October. He compared GPT-6 Astra and Claude Fable 5.1 with Inception v3, a CNN that Google published in 2015, fine-tuned on public skin images. On 200 biopsy-proven lesions, the CNN named the right diagnosis first 62% of the time, against 42% for both LLMs. The models failed in different ways: GPT-6 Astra answered basal cell carcinoma for 75 lesions, while the CNN, at its preset threshold, sent home 18 of 30 melanomas against 3 and 4 for the LLMs. Lynch lists the limits: a small sample, one run per model, and a dataset public since 2025. Why it matters: on a narrow visual task, a small trained model can still beat a general chatbot, a useful point when clients want an LLM for everything. Source

Glossary

  • Wiz: a cloud security company; Wiz Research is its threat research team.
  • Infostealer: malware that copies saved passwords, cookies and access tokens from an infected computer and sends them to an attacker.
  • LiveCodeBench: a benchmark that scores AI models on competitive programming problems.
  • Shadow AI: AI tools that employees use for work without their employer’s approval.
  • Open-weight model: a model whose trained parameters are published, so anyone can download it and run it on their own servers.
  • GPT-6 Astra: a model from OpenAI’s GPT-6 family.
  • CNN: a convolutional neural network, an older type of AI model built to recognise patterns in images.