Sam's News — gpt — 2026-10-14¶
Security¶
8 Researchers weaponize public OpenAI model to build zero-day Chrome exploit in six days¶
XBOW researchers used OpenAI's publicly available GPT-5.6 Sol model to chain three Chrome zero-days into a full exploit in six days, costing ~$40,000 in AI charges. Google awarded the maximum $250,000 bounty after confirming and fixing all three vulnerabilities.
- Exploit built in 6 days (August 2–7, 2026) using public GPT model
- AI charges cost approximately $40,000 per usage logs
- Google awarded maximum $250,000 bug bounty
- Vulnerabilities: CVE-2026-85045 (V8), CVE-2026-78952 (WebP decoder), and one unpatched Chrome component
- Before LLMs, comparable exploit required ~4 engineers over 6–9 months
Sources: The Hacker News AI Web Searched
7.5 OpenAI engineer charged after surveillance device found in executive conference room¶
OpenAI engineer Abdulrahman Alfozan, 35, a Saudi national employed since July 2025, was charged with attempting to intercept oral communications after FBI discovered a hidden recording device in the executive conference room. The device, concealed in toilet paper and plugged into a television on September 28, may have been modified to prevent data transfer.
- Abdulrahman Alfozan, Saudi national, worked at OpenAI since July 2025
- Seeed XIAO ESP32-S3 Sense recording device found in TV on September 30
- Second device discovered in Alfozan's bag; microSD card contents encrypted or unrecoverable
- CCTV placed him near room multiple dates despite lacking conference room access
- Arrested October 9; initial court appearance October 13; no definitive evidence data captured
Sources: The Register AI Web Searched, Türkiye Today RSS, The Federal RSS
6.5 OpenAI Adds Text Watermarking to ChatGPT and Codex¶
OpenAI is implementing text watermarking features in ChatGPT and Codex to detect AI-generated content.
Sources: The Verge Web Search
AI¶
7.5 Microsoft Releases AI Model 35 Times Faster Than OpenAI's GPT-6¶
Microsoft released Microsoft-Decision-1, a decision model 35 times faster than OpenAI's GPT-6 Sol on structured tasks. Built by fine-tuning Alibaba's Qwen3.5-9B, it achieved highest accuracy across 36 benchmarks (150,000 blind questions) and costs $0.042 per million input tokens with free output.
- 35× faster than GPT-6 Sol on structured decision tasks; 2.5× faster than H2O-Lightning-4B v1.1
- Highest accuracy across 36 benchmarks, 150,000 blind test questions
- 32,000-token context window; text input only
- $0.042 per million input tokens; output tokens free
- Xbox Research: competitive with GPT-6 Sol at 14× faster speed, 1/200th cost; Copilot: competitive with GPT5.6 Luna at 100× faster speed
Sources: Microsoft Official Blog AI Web Searched, The Next Web AI Web Searched, EdTech Innovation Hub AI Web Searched, Barchart.com Web Search
7 Google launches Gemini 4 Argon AI model globally¶
Google launched Gemini 4 Argon globally with published specifications and API pricing. Priced at $2 per million input tokens and $4 per million output tokens, the model positions Gemini 4 Argon as a competitor to rival AI models.
- Gemini 4 Argon pricing: $2 per million input tokens, $4 per million output tokens
- Global launch with published specifications
- Described as 'uncensored cyber defense' offering
- Represents generational leap in language model capability
- Gemini 4 Argon released October 13 after 13-day restricted cybersecurity defender access
- Introductory pricing: $2 per million input tokens, $10 per million output tokens through November 30, then doubles
- 1 million combined input/output token limit with continuation protocol; 95% discount on cached input
- 100,000+ Google employees already use Argon monthly
- Vals AI ranks Argon first (68.90% vs Claude Sonnet 5.5's 67.04%); Artificial Analysis: tied at 53 with GPT-6 Astra and Claude Fable 5.1
Sources: Google News AI Web Searched, Tech AU RSS, Yellow.com AI Web Searched
7 ChatGPT and Codex retire to GPT-6.1 Sol as GPT-5.5 reaches end of life¶
OpenAI is transitioning ChatGPT and Codex services from GPT-5.5 to GPT-6.1 Sol. Codex has expanded from code-only to general computer work, file handling, web research, and connected applications, now integrated into ChatGPT desktop app.
- GPT-5.5 reaching end of life; transition to GPT-6.1 Sol
- Codex expanded beyond code to general computer work, file handling, web research, and connected applications
- Codex integrated into ChatGPT desktop app; also available via terminal CLI, IDE extensions, or cloud service
- Supports bounded agentic tasks: repository inspection, bug fixes, migrations, test work
Sources: Kingy AI - OpenAI Codex Guide (2026) AI Web Searched, https://tech-insider.org/ RSS
6.5 Google releases Gemini 4 Argon, competing with GPT-6 and Claude¶
Google has released Gemini 4 Argon as a public offering to compete with OpenAI and Anthropic's latest large language models.
Sources: Yellow.com RSS
6 Higgsfield Extends AI Video Generation to Over 10 Minutes¶
Higgsfield's technology now generates AI video content exceeding 10 minutes in length, a substantial improvement from previous 1-minute limitations.
Sources: OpenAI Web Search
AI Ethics¶
6.5 OpenAI Accused of Mining Children's ChatGPT Data for Training¶
OpenAI faces accusations of using data from children's ChatGPT interactions to train its large language models without proper consent.
Sources: Law360 RSS
Policy¶
6 EU's AI regulation enters agentic phase¶
The European Union's regulatory framework for artificial intelligence progresses toward addressing autonomous AI agents.
Sources: OpenAI RSS