Top line
Chrome 157 update fixes 83 security bugs, five Critical
Google pushed Chrome 157.0.8103.78/.79 for Windows and Mac and 157.0.8103.78 for Linux on Wednesday at 2:02 p.m. CT, rolling out "over the coming days/weeks." The release has 83 security fixes: by our count of the post, 5 Critical (CVE-2026-126885, -126895, -126879, -126846 and -126911, mostly use-after-free bugs) and 78 High. Google doesn't list any as exploited. At least 11 fixes credit Anthropic researchers or OpenAI's Codex Security, a sign that AI-assisted bug hunting is now feeding browser patch volume. Extended Stable moved to 156.0.8078.100. Microsoft Edge's desktop build was still 157.0.4341.46 at our cutoff, so it hasn't picked these fixes up yet.
Why it matters: Push Chrome now instead of waiting for the staged rollout, and watch for a matching Edge build. None of these is known to be exploited yet, but the volume is unusually high.
Sources: Chrome Releases · Edge security notes
Gemini Decision 1: bulk yes/no and rubric checks for $0.10 per million tokens
Google launched the Gemini Decisions API, powered by a new gemini-decision-1 model, on Wednesday at noon CT. One request can ask up to 1,000 yes/no, multiple-choice or rubric-scoring questions about a single input of up to 262,144 tokens of text, images, audio, video or PDF. Each answer comes with a confidence score from 0 to 1, and the model generates zero output tokens. Shared input is billed once per request: $0.10 per million input tokens on the paid tier, free on the free tier. The Batch API isn't supported, and free-tier data is used to improve Google's products. It runs on the v1beta API, and the docs don't say whether it's preview or GA.
Why it matters: Content moderation, document triage, eval grading and routing checks that now run through a full chat model may cost far less here. Keep sensitive data on the paid tier.
Sources: Google announcement · Model page · Pricing
Researchers chain a prompt injection toward a backdoored Codex release
Security firm Depi published a write-up dated Nov. 11 showing how a prompt injection in a GitHub issue could hijack the AI issue-labeler workflow in OpenAI's Codex repo. From there they chained a sandbox escape and GitHub Actions cache poisoning into a path to publishing a backdoored @openai/codex npm package. Per the researchers, OpenAI "deployed a fix right away" and paid an initial $1,000 bounty. Bugcrowd lists the report as P1 and resolved. OpenAI hasn't commented publicly and there's no CVE. Nothing suggests a malicious release actually shipped.
Why it matters: Any repo that lets an AI agent read untrusted issues inside a CI workflow with write tokens or a shared cache faces the same risk. Check yours now.
Sources: Depi write-up · Bugcrowd disclosure
What to do
- Patch Barracuda appliances by Nov. 13. CISA's deadline for CVE-2026-116606 and CVE-2026-111560 is Friday, and the catalog had no new additions over the Veterans Day holiday. Barracuda advisory
- Update LiteLLM proxies that use Azure models without an API key. GHSA-2rwg-2wh4-7c4f (High) lets a logged-in low-privilege user pull secrets, including the master key, through Azure Entra parameters in the request body. It's fixed in 1.106.9, 1.108.6, 1.109.1 and 1.110.0-rc.2. GitHub advisory
- Move off GPT-5 and o3 before Dec. 11. OpenAI removes gpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3 and o3-pro from the API that day, fine-tunes included. Recommended replacements are gpt-5.6-sol, -terra and -luna. Notice went out in June, and OpenRouter now flags the same date. OpenAI deprecations
- Check for eMudhra (emSign) certificate chains before Nov. 16, when Windows, Edge and WebView2 stop trusting them by default. MSRC blog
Also notable
- OpenAI: Workspace Agent runs were down for nearly 10 hours. New runs couldn't start from 4:50 p.m. CT Wednesday to 2:32 a.m. CT Thursday, with no cause given. OpenAI status
- GitHub: another critical incident. Actions, Pages, Copilot and Webhooks were degraded from 2:48 to 3:21 p.m. CT Wednesday, with a root-cause write-up promised. Write-ups for earlier November incidents are still missing. GitHub status
- Gemini Managed Agents can run in your own sandbox. The sandbox connects out over HTTPS through the open-source gemini-api CLI, and hosted-only options return errors there. Google AI Studio docs
- Google Agent Platform: per-user
safety_identifieris GA for Gemini and Veo calls, so multi-tenant apps can tie abuse to one end user instead of the whole project. Release notes - Baidu World: a token marketplace and new ERNIE app models. Baidu says its Token Exchange passed 50 million API calls a day in its first month, and it claims new Qianfan models cut inference costs by 30% or more. No prices were given. Yicai Global · Baidu
- Qwen Code 1.0 Preview (catch-up). Alibaba's open-source terminal coding agent reached 1.0 preview under Apache-2.0 on Nov. 10. GitHub release
- AWS: Web Search API for Bedrock AgentCore. One IAM-signed call returns ranked web results for agents. Pricing isn't stated. AWS blog
Sources
Chrome Releases · Edge security notes · Google Decisions API announcement · Gemini Decision 1 model page · Gemini pricing · Depi write-up · Bugcrowd disclosure · Barracuda advisory · LiteLLM advisory · OpenAI deprecations · MSRC eMudhra blog · OpenAI status · GitHub status · Gemini agent environment docs · Agent Platform release notes · Yicai Global · Baidu · Qwen Code release · AWS AgentCore blog