AI Weekly for IT Pros: November 7–13, 2026

The biggest story for IT teams this week was Anthropic’s release of Claude Fable 5.2 and its restricted twin Mythos 5.2 on November 9: one model, two access tiers, cheaper API pricing and new data-governance options that matter to anyone running AI in a regulated environment. Meanwhile Google and Meta both shipped cheaper “workhorse” models for agents, and November’s Patch Tuesday put a critical Azure OpenAI flaw and an exploited Outlook bug on your list.

Top stories

• Anthropic ships Claude Fable 5.2 and Mythos 5.2. Fable 5.2 (claude-fable-5-2) costs $8/$40 per million input/output tokens, with a fast mode at $16/$80, and is available on the Claude Platform, AWS, Google Cloud and Microsoft Azure. Mythos 5.2 is the same model without the cyber and biology blocks and is limited to vetted programmes. Admins should note the new Enterprise Frontier Safeguards private beta, which runs safety monitoring inside the customer’s own infrastructure and works with zero data retention. Other customers keep 30-day retention for safety monitoring.
• Google launches Gemini 4 Hydrogen as its fast, low-cost model. Released November 12, Hydrogen takes over the “Flash” role in the Gemini 4 family and is available in the Gemini API, Google AI Studio, Android Studio, Gemini for business and Google Search for all users. For teams running agent pipelines, the useful pattern is pairing it with a bigger model: Google’s Antigravity Orchestrator hands routine steps from Gemini 4 Carbon to Hydrogen.
• Meta releases Muse Ember for repetitive agent work. Announced November 9 at $0.10 per million input tokens and $0.50 per million output, with a 1M-token context, Ember targets tool calls, routing, data extraction and long-document parsing. It’s available through the Meta Model API, Google Cloud and Oracle Cloud. It’s another option to benchmark if token bills from high-volume automations are growing.
• Claude joins your video calls. Anthropic’s Claude Meetings (beta, Team and Enterprise plans, Google Meet at launch) adds Claude as a visible participant that drafts plans and pulls data from connected apps during the call. By default it won’t join meetings with external guests unless an admin allows it, and it leaves if one joins. Review this setting before users start inviting it.
• Patch Tuesday: critical Azure OpenAI bug, exploited Outlook flaw. Microsoft’s November 10 release includes CVE-2026-119990, a CVSS 9.8 Azure OpenAI elevation-of-privilege bug (Microsoft now says no customer action is needed), and CVE-2026-124194, an Outlook spoofing bug that is under active attack and leaks NTLMv2 hashes when a user opens a crafted message file. Patch Outlook first.
• Fireworks AI breach widens. The inference provider’s November 9 update says attackers also reached a subset of customer datasets, incomplete model weights and, for some accounts, secrets, after a third-party cloud provider incident. If you use Fireworks, check for a notification email and rotate any secrets stored with the service as a precaution.

AI in Microsoft & Azure

• o-series models retire in Foundry next week. Microsoft’s retirement schedule lists o1, o1-pro, o3, o3-pro and o3-deep-research (replacement: gpt-5.6-sol), plus o3-mini and o4-mini (replacement: gpt-5.6-terra), for November 19. codex-mini and claude-haiku-4-5 go on November 15 with no listed replacement. Retired models return 410 Gone, and Provisioned deployments are not auto-upgraded.
• Azure OpenAI and MCP Gateway CVEs. Besides the critical Azure OpenAI fix, November’s release covers CVE-2026-109492 (Microsoft MCP Gateway elevation of privilege, rated Important) and CVE-2026-83953 (GitHub Copilot for JetBrains information disclosure). Check Microsoft’s advisories for any action on self-managed components.
• Microsoft signs the Paris Call for AI Safety and Security. Microsoft, OpenAI, Anthropic and Mistral AI are among the ten companies that signed on November 9. The commitments include shutdown mechanisms for frontier models and stronger incident reporting. Google and Meta did not sign.

Tools for admins and developers

• VS Code 1.146 deprecates the Local agent harness. The Copilot harness is now the path forward. The Local harness will be removed in February 2027, along with the enterprise setting vscode.temporaryDefaultToLocalHarness. If you pinned users to the Local harness, plan the switch now.
• New VS Code controls worth a look. 1.146 adds sandbox credential masking (chat.agent.sandbox.credentials.envVars), narrower default permission prompts, a communication.support.url managed setting that adds “Contact Admin” links to policy blocks, and support for proxy settings (including PAC and Kerberos) in Copilot sessions.
• Claude Fable 5.2 fast mode reaches Microsoft Foundry. Fast mode (up to 2.5x faster output) is available in Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry (Anthropic-hosted). Anthropic still recommends starting most coding work on Opus 5.5 and moving up only when needed. That’s a useful default for controlling cost.
• Gemini 4 Hydrogen for coding agents. Google positions Hydrogen for repository coding, debugging and web development, with four thinking levels. It’s available to developers now in AI Studio and the Gemini API, and in Antigravity for paid AI Pro and Ultra subscribers.

AI security and compliance

• Microsoft Research shows how to backdoor an AI agent by editing under 10 neurons. The “Lazarus” technique edits an open-weight model’s dormant neurons without retraining, so the model passes evals until a trigger (such as the date rolling into 2027) makes the agent download and run attacker code. It needs write access to the weights. Defences: verify model provenance against the original weights, lock down model storage and pipelines, limit agent permissions and require review for high-impact actions.
• Attackers abuse an open-source AI pentest platform against AWS. Arctic Wolf found attackers running self-hosted AntiFox instances to plant web shells on IIS servers and harvest AWS secrets with stolen credentials. AWS recommends least privilege, IMDSv2, no long-term access keys in app configs, and GuardDuty. Also watch for successful API calls hidden in bursts of AccessDenied errors.
• Know which model tier your users get. Fable 5.2 blocks high-risk cyber requests and routes them to an older model, while Mythos 5.2 needs Cyber Verification. Claude Enterprise customers get Claude Security, powered by Mythos 5.2, without that verification. Security teams should check which tier fits their red-team and code-scanning work.

Try this week

Before the November 15 and 19 Foundry retirements, list every model deployment on your Azure OpenAI / Foundry resource with its version, SKU and upgrade setting. Look for retiring models, deployments set to NoAutoUpgrade (they stop working at retirement) and Provisioned SKUs (they must be migrated manually). Replace MY-RG and MY-ACCOUNT with your own names:

az cognitiveservices account deployment list \
  --resource-group MY-RG --name MY-ACCOUNT \
  --query "[].{deployment:name, model:properties.model.name, version:properties.model.version, upgrade:properties.versionUpgradeOption, sku:sku.name}" \
  --output table

Re-test anything that will be upgraded. A newer model can change token usage and costs, so run your evals before the date, not after.

Sources

• Anthropic: Claude Fable 5.2 and Mythos 5.2
• Google: Gemini 4 Hydrogen for developers
• 9to5Google: Gemini 4 Hydrogen launch
• Digital Trends: Meta launches Muse Ember
• TechCrunch: Anthropic launches Claude Meetings
• Zero Day Initiative: November 2026 security update review
• Fireworks AI: security incident update
• Microsoft Learn: Foundry model retirement schedule
• Microsoft Learn: model deprecations and retirements
• Microsoft Learn: az cognitiveservices account deployment
• VS Code 1.146 release notes
• France Diplomacy: Paris Call for AI Safety and Security
• Microsoft Research: backdooring AI agents via dormant neurons
• BleepingComputer: AntiFox AI agents in AWS credential-theft attacks

Questions about using AI in your IT environment? Get in touch.