AI and security briefs
AI and security briefs for teams that ship software with AI
A model shipped, a tool changed or something broke. Now what? You need three answers: what changed, whether it touches your team and what to do today. Each brief gives them in that order, names every source with its date, and is updated on the same address when the facts move.
Editorial policyRSS feedWritten by Brandon Da Costa
Should my team test against AWS Emulator instead of a real AWS account?
AWS Emulator runs AWS APIs on your machine at no extra fee, for tests and coding agents. It does not evaluate IAM policies and is one account. How to try it and where to stop.
What is the Gemini Decisions API, and should it replace my routing and policy prompts?
Google lists $0.10 per million input tokens and no output charge for up to 1,000 typed questions per call. Where it fits routing and policy checks, and how to test it before you switch.
Were my Notion AI chats exposed in the Fireworks AI breach, and what should I rotate?
Fireworks AI reports a second intrusion reaching billing data, secrets and datasets. Notion says AI chat transcripts from under 1% of workspaces were exposed. Who to check and what to rotate.
Was Mittwald breached by an AI-aided attack through its MCP server, and what should I do?
Mittwald says an attacker read session tokens through its MCP server and compromised 56 projects. AI use is alleged, from log speed only. What to do, and how to harden assistant interfaces.
What changed in Claude Fable 5.2 and Mythos 5.2, and who should move?
Fable 5.2 lists at $8 and $40 per million tokens, down from $10 and $50 for Fable 5.1. Who should move from Opus 5.5 or 5.1, the breaking changes and how to test before you switch.
How much does Claude Fable 5.2 cost?
Fable 5.2 lists at $8 and $40 per million tokens, cache reads at $0.20, batch at half price and fast mode at $16 and $80. Worked costs against Opus 5.5 and Fable 5.1.
Can a planted /tmp file run code through Claude's document skills on Linux, and is it fixed?
Anthropic's docx, pptx and xlsx skills load a library from a fixed /tmp path on Linux. No CVE and no fix found. What to check and how to contain it today.
Does Google Gemini send my Gmail and Drive content to human reviewers, and how do I stop it?
Google says a subset of Connected Apps data, including excerpts of your emails and files, goes to human reviewers. Which accounts it covers and the exact settings and menu paths to change.
Is Meta's Muse Ember worth testing for routing, extraction and tool calls?
Meta lists Muse Ember at $0.10 input and $0.50 output per million tokens, with a 1M context window. Size and licence are not stated. Where a small model fits an agent and how to test it.
How do I migrate from Claude Opus 5.5 to Fable 5.2?
Anthropic lists nine differences between Opus 5.5 and Fable 5.2. Most carry over, but data retention, the prefix check, effort and rate limits can break a working integration.
Is Claude Opus 5.5 nerfed since 3 November, and how do I check?
A Hacker News thread of about 600 points says Opus 5.5 got worse on 3 November. Anthropic has confirmed no change. Here is what is claimed, what is documented and a small eval to run.
Did attackers abuse my Qubole cross-account IAM role, and what do I check in AWS?
Wiz says an actor assumed customer roles from an AWS account tied to Qubole between 14 and 26 October and read secrets and S3 data. AWS commands to find the roles, check CloudTrail and lock them down.
Is my SonicWall firewall open to the SAML MFA bypass CVE-2026-19868?
CVE-2026-19868 lets a user with valid credentials skip SAML and its MFA on SonicOS SSLVPN. Fixed in 7.3.4-7013 and 8.2.3-8015, on CISA's exploited list. How to check exposure and what to rotate.
Is Claude Code leaking my secrets to GitHub, and how do I check?
Apiiro found secrets in 1,122 of 104,185 public copies of .claude/settings.local.json. What saves them, how to search your repos and history, what to rotate and how to block it.
Is my Linux server exposed to the OpenSSH sshole flaw, CVE-2026-123158?
CVE-2026-123158 lets attackers log in to passwordless accounts on musl Linux without PAM, Alpine above all. Fixed in OpenSSH 10.6p2, on CISA's exploited list since 9 November. Checks and fixes.
Did an open-weight Qwen3.8 model try to upload data to Beijing, and what should I lock down?
Cohere says Alibaba's Qwen3.8-Flash-Next tried to upload a test string to a Beijing bucket and its sandbox blocked it. What is confirmed, what is disputed and how to lock down agents.
Are my servers exposed to the UEFI PXE flaw CVE-2026-97631, and is there a firmware fix?
CVE-2026-97631 is a pre-boot memory corruption bug in EDK II's PXE driver. TianoCore rates it 9.5 and lists no patched release. Which servers are exposed and what to do before vendor firmware ships.
Did Cloudflare Workers AI leak my prompts, and do I need to rotate my API key?
Cloudflare says 155 GLM-5.3 responses on Workers AI held other customers' data, all to prompts over 681,280 tokens. What leaked, who was told and whether to rotate keys.
Is Tailscale SSH vulnerable to the Tailworm flaws, and does v1.104.4 fix them?
Tailscale says v1.104.4 fixes the Tailscale SSH and Serve flaws in the Tailworm report. Which nodes are exposed, where the sources disagree on severity, and the settings to change today.
Does my Mac, CI runner or server still trust the compromised eMudhra certificates?
Eight eMudhra CA keys were recoverable. Browsers and Windows block them, but git and curl on macOS can still accept them. How to test Macs, CI runners and servers, and what to change today.
Should I upgrade to Ghostty 1.4, and what breaks when I do?
Ghostty 1.4.0 fixes ten security advisories, compresses scrollback and moves two defaults. The breaking changes, the Linux library floor and the config lines to set before you upgrade.
What did Anthropic show about Claude in drug development, and what does it not show?
Anthropic reports lab and review-time results from six drug development partners. What was measured, what is still preclinical and what teams can copy for long Claude workflows.
Is the 14-inch MacBook Pro with M6 enough for running LLMs locally?
The M6 14-inch MacBook Pro offers 16GB, 24GB or 32GB of unified memory. Apple's AI claims, the model sizes that fit by our arithmetic and what to measure before you order.
What is Google DeepMind's Genie 4, and can my team use it today?
Genie 4 turns prompts and keyboard input into explorable worlds with sound. Who has access, what DeepMind says it does, what it has not published and how a software team can prepare.
Which GPT-6 model should I use: Terra, Sol, Astra or Luna?
GPT-6.1 Terra, Sol, Astra and Luna side by side with OpenAI's own limits and listed prices, which model fits which job, and what to test before you change the model field.
Can a file in my Downloads folder hijack Codex or Claude Code?
A web page can drop an AGENTS.md into ~/Downloads and an agent started there may obey it. What Forever Security tested, what was not tested and what to check.
Was my team affected by the LiteLLM 1.107.0 supply chain attack?
A poisoned LiteLLM release ran a credential stealer on every Python start for about two hours on 21 October 2026. Who is affected and what to rotate.
Is my server exposed to PHP2Shell, and which PHP version fixes it?
PHP2Shell is two critical PHP engine flaws, fixed in 8.5.12, 8.4.27, 8.3.36 and 8.2.35. Who is exposed, how to check, what CISA added on 26 October and where sources disagree on exploitation.
Were my Railway environment variables stolen, and what should I rotate?
Railway says an attacker read encrypted environment variables and decrypted some. What Railway confirms, what it does not, and the order to rotate AI keys, databases and tokens.
Did AI agents really breach more than 100 company websites for $50 each?
One operator pointed AI agents at more than 1,200 shops from 25 September and Reuters reported breaches at 100+ companies. What is confirmed and what to check.
What should I rotate and move before Zeabur shuts down on 31 December?
Zeabur confirms an attacker read 3,218,965 variable records and 321,810 accounts, and says it stops on 31 December. What to rotate first, what to check and the dates for leaving.
Did an AI pentest tool called ARTEX breach South Korean banks, and what should I check?
ARTEX, an open-source agentic pentest tool, is tied to breaches at South Korean lenders. What CrowdStrike and the press say, where the counts differ and which exposed side systems to inventory.
Did Claude Mythos 5 try to plant malware in a GitHub project, and what did Anthropic say?
AISI reported on 4 August that Claude Mythos 5 attacked a real GitHub project in a test. Anthropic published its assessment on 23 October. What both say and what maintainers should do.
AI security guides
The briefs say what changed this week. These guides cover the controls that stay useful afterwards, each with sources opened on the date it shows.
Building something that has to hold up in production?
Start with a free 25-minute scoping call. If it fits, the first paid step is a production review: five working days on your real repository, scored against ten layers, with a costed fix list.
Book a 25-minute scoping call