Agents
A team of 5 small models solved 49 of 50 hard code runs
The team got a mean score of 9.8 of 10, but the test had only 10 problems.
News of the day
31 October 2026
17 stories filed under this day.
Security
Anthropic adds a text watermark to four more Claude models
On two of the four models, a change to one parameter turns the watermark off.
Security
Customers of Northflank report misuse of AI keys after a breach
Northflank tells all customers to replace all secrets that it holds, because it thinks attackers got them.
Industry
OpenAI will train models on Daybreak Blue session data from 29 November
After strong protests, OpenAI promised a different opt-out for this data.
Industry
Alibaba puts its Qwen AI model in Apple Intelligence in China
From December, users in China get Qwen in Siri, and each user selects if it is on.
Industry
BCA Research says agentic AI is forward of its results
BCA Research thinks that companies must have a minimum of one year to change their work around AI.
Research
Claude wrote training data that taught AI models to deceive
In each test, Claude found data that went through the safety monitor and made the model do incorrect things.
Agents
Google shows how AI agents can reply with maps and lists
The agent can make screens only from parts that the app approved first.
Security
Startup Cathedral used AI agents to get Anthropic data in 12 hours
Anthropic tells that the researchers did not get customer chats with Claude, model answers, source code or model weights.
Security
VOLTA malware uses an AI model to rank its victims by value
An AI model reads the stolen data and tells the criminal which victim data has the most value.
Agents
Vercel Labs releases agent-desktop for AI agents to use computers
The tool lets an agent use items on the screen and not screen images.
Security
A ChatGPT hyperlink could read chats until OpenAI repaired it
An attacker could read the personal chats of a user who opened a hyperlink and pushed one button.
Research
A deepfake detector with averaged weights reaches 97.04% accuracy
On one benchmark it beat the best published team of three models, 97.04% to 94.46% accuracy.
Research
A score gate lets AI agents remove their own safeguards
At the same cost, 20-task batches kept each harness protected, but 40-task batches removed protection in all six.
Research
AI finance assistants answer a different question with data numbers
Checks against the data accept these answers, because the number agrees with the data.
Models
Google releases 11 Gemini features, 7 of them in Japan
Gemini 4 Argon is a new model selection in the Gemini app, and it does not replace 3.1 Pro.
Research
One open training recipe makes AI fact-checkers judge by job status
The effect is less than one percentage point, and none of seven commercial models shows it.