Happy Wednesday. A new piece of malware discovered by Cisco Talos runs its own attack strategy by polling four commercial AI models and acting on the majority vote, with no attacker sitting behind a keyboard.
On the same day, Anthropic cut cache-read costs 60% with a new Opus release, and OpenAI answered 90 minutes later with two models priced at half what their predecessors cost. If you run agentic pipelines, the math on your API bills changed yesterday.
SECURITY & PRIVACY
Cisco Talos Finds Malware That Runs Itself With an AI Voting Panel
Cisco Talos disclosed CLOSEDQUORUM, a Windows credential stealer that queries Google Gemini, DeepSeek, Alibaba Qwen, and Mistral every five to 15 minutes, executes the majority-vote result, and needs no human operator after deployment. The malware targets saved browser passwords, Windows credentials, and crypto wallet data from MetaMask and Exodus. Talos found no confirmed in-the-wild deployments, and the sample they analyzed shipped with placeholder API keys and missing lateral-movement code.
The detection problem is the part security teams should sit with. CLOSEDQUORUM phones home to the same API endpoints used by thousands of legitimate apps, so there is no rogue server to block and no unusual outbound destination to flag. Talos released an open-source detection toolkit called CAIRN that scans file metadata for embedded prompts and model API strings rather than requiring execution. That is worth evaluating now. Talos researcher Ryan Fetterman noted that this architecture can be applied to other attacker objectives, meaning the voting-panel design is a template, not a one-off.
ARTIFICIAL INTELLIGENCE
Claude Opus 5.5 Beats GPT-6 Astra on Coding Benchmarks at 40% of the Cost
Anthropic released Claude Opus 5.5 on Sept. 22 at $4 per million input tokens and $20 per million output tokens. Cache-read costs dropped 60% to $0.20 per million tokens. In early testing, the model completed a 680,000-line code migration in under a day and a 200,000-line codebase audit in under three hours, a task that took Opus 5 more than 20 hours and 2.5 times as many tokens.
The cache-read cut is the number that matters most for teams running multi-step agentic workflows, because those reads stack up fast and the discount requires no code changes to realize. The benchmark results against Fable 5.1 and GPT-6 Astra give procurement teams a concrete reason to retest current production models. Deloitte's CIO reported Opus 5.5 caught 72% of known bugs at its lowest effort setting, compared with 56% for Opus 5 at high effort. That kind of accuracy-at-lower-cost combination is what rewrites vendor selection decisions.
ARTIFICIAL INTELLIGENCE
OpenAI Makes 50% Price Cuts on Sol and Luna Permanent, Overhauls Caching
OpenAI released GPT-6 Sol and GPT-6 Luna on Sept. 22, cutting prices at least 50% from their GPT-5.6 predecessors. Sol lands at $2 per million input tokens and $10 per million output tokens. Luna's output price fell 58% to $0.50 per million tokens. OpenAI said the cuts are permanent and stem from caching and inference efficiency gains. Cached input-token reads now carry a 90% discount, and GitHub reported the overhaul cut fresh token processing by more than 50% across billions of requests.
The timing matters as much as the pricing. Anthropic published Opus 5.5 roughly 90 minutes before OpenAI's announcement, and OpenAI's DevDay follows on Sept. 29. For developers, the back-to-back drops mean the cost basis for production AI applications is falling faster than most budgets anticipated at the start of this year. Sol's AutomationBench score of 33.2% at $0.27 per task, against Claude Opus 5's 26.9% at more than 11 times that cost, gives enterprise buyers a direct data point for mid-tier model selection.
Created by the robots at The SaaS Sentinel