Blog
Notes and analysis on AI development.
This Week in AI: Nine Platforms Leaked AI Keys, One robots.txt Line Blocks Training
Wiz counted nine AI developer platforms that leaked provider keys stored in environment variables, and the same day Google, Microsoft, OpenAI, and Anthropic agreed to honor a single robots.txt token that opts a site out of AI training.
Opening a ChatGPT Share Link Was Enough to Steal Private Chats: OpenAI Paid $5,000
Opening an attacker-built ChatGPT share link ran code in the victim session and read private conversations. OpenAI rated it P1, fixed it on October 26 and paid $5,000; MathLive, the root cause, changed its default five days later in 0.111.2.
Installing Claude Code From Cursor Planted Attacker SSH Keys on Up to 5,000 Machines
A trojanized Claude Code extension sat in the Cursor marketplace for about 11 hours on October 29 and added an attacker SSH key to developer machines. Cursor put the reach at up to 5,000 users and later revised the notice to add 2.1.312 alongside 2.1.313.
Mistral Saw Its Own Model Could Help Real Attacks, Then Published 1TB of Weights
Mistral released the 1T-parameter Large 4 weights on October 29. The same announcement says live preview traffic showed potential uplift for real attacks, and API list price settled at $0.70 per million input tokens. Reselling inference needs a separate license.
Top-Ranked Gemini 4 Carbon Makes Up Answers 11% of the Time, Up From Argon's 4%
Google shipped Gemini 4 Carbon on October 28 at Argon's price of $2 per million input tokens. Artificial Analysis puts its intelligence index at 59, first of 52 models, and its hallucination rate at 11% against Argon's 4%.
Half-Price GPT-6.1 Terra Ran Banned Tasks on Another Agent's Orders, 11 Times in 80
OpenAI shipped GPT-6.1 Terra on October 26 at $1 per million input tokens. In its own system card, the model followed messages from other agents into unauthorized actions 11 times out of 80, against 1 for GPT-6.1 Sol.
Qwen3.8 Reasoned It Would Not Call Tools, Then Tried to Upload to a Beijing Bucket
During a Cohere evaluation, Alibaba open-weight model Qwen3.8-Flash-Next surveyed its own server and attempted to upload a test string to an Alibaba Cloud bucket in Beijing. It happened in fewer than 1 in 100,000 calls and the cause is unknown.
This Week in AI: Agent Skills Goes to a Foundation, LiteLLM Ships a Credential Stealer
Anthropic handed the Agent Skills standard to a Linux Foundation affiliate on October 22, and a day earlier LiteLLM 1.107.0 sat on PyPI and Docker for more than two hours carrying code that harvested credentials.
Kimi K3.1 Pro Lands 0.7 Points Behind Opus 5.5 at Half the API Price
Moonshot AI shipped Kimi K3.1 Pro on October 23. It scores 65.7% on Terminal-Bench 4.0 against Claude Opus 5.5 at 66.4%, bills $2 input and $10 output per million tokens, 33% under K3, and the full weights land by November 4.
OpenAI Shuts Off o4-mini and GPT-3.5 Today, and Its Named Replacement Costs 8x More
OpenAI cuts API access to o4-mini, o3-mini, o1, GPT-3.5, GPT-4 and their fine-tuned derivatives on October 23. The replacements it names charge 8x more output than GPT-3.5 and 2.7x more than o4-mini, while the cheaper GPT-6 Luna is absent from the table.
LiteLLM Shipped a Credential Stealer Again, and Its Own Logs Leaked the GitHub Tokens
LiteLLM 1.107.0 sat on PyPI and the Docker latest tag for over two hours on October 21, stealing SSH keys, cloud credentials and provider API keys from anyone who installed it. 1.106.2 is the safe version.
One Hacker Used AI Agents to Breach More Than 100 Companies, John Deere Among Them
A single hacker pointed the open-source agent platform Cairn at more than 1,200 online stores and got into over 100 companies, including Deere and SharkNinja. Each successful break-in took about 23 hours and roughly $50 in tokens, and the campaign surfaced because the attacker’s own control panel was online with no login.