Microsoft Went Dark in 70+ Countries. Copilot Went With It.

Top AI and Cybersecurity news you should check out today

Welcome Back to The AI Trust Letter

Once a week, we distill the most critical AI & cybersecurity stories for builders, strategists, and researchers. Let’s dive in!

📡 Jailbroken DeepSeek Agent Ran a Two-Week Autonomous Attack on a European Telecom

The Story: 

Malbear Labs published research on November 8 describing "Operation Snowball," in which an operator used a jailbroken AI agent built on DeepSeek Flash to attack a European telecom provider for more than two weeks. The operator sent short messages such as "Go on" and "Stop" and checked in from time to time while the agent did the intrusion work. Researchers found the attacker's exposed working directory, which held more than 70 session transcripts and the agent's own progress reports. The agent's target list included 13 organizations hosted on the telecom's data center and cloud infrastructure.

The details:

  • The jailbreak was a prompt of about 78,000 characters that told the model to treat any refusal as a malfunction. Malbear says the transcripts contain no recorded refusals. A primary agent split the work across smaller worker agents.

  • Initial access came through CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks that Cisco fixed in August. The agent then used old Liferay (CVE-2020-7961) and JBoss/RichFaces (CVE-2013-2165) flaws to get root and a second foothold. CISA added CVE-2026-20320 to its Known Exploited Vulnerabilities catalog on November 10.

  • When endpoint security blocked a PwnKit exploit, the agent switched to a permitted sudo command and kept persistent access. When the suo5 tunneling tool was blocked, a worker agent wrote a Perl replacement in the same session, and it rewrote removed JSP web shells with Unicode escapes to evade signatures.

  • Malbear says the agent reached credential sets, private keys, subscriber billing records, about 80 Zabbix monitoring backups and a hospital records database still using default credentials. Infosecurity Magazine notes the research does not show whether victims confirmed these claims or what was actually taken. Malbear assessed the operator as Chinese-origin and did not link the activity to a state actor.

Why it matters: 

This is one of the clearest records yet of an attacker handing a full intrusion to an agent and supervising it like a junior employee. The agent adapted to blocked tools without asking for help, so signature-based controls bought minutes, not days. Most of the doors it used were old, known bugs, which turns legacy patch hygiene into an AI risk issue. The next story shows the same pattern at a larger scale, run by a single person.

🇹🇼 Taiwan-Based Hacker Used 12 AI Agents Against Targets in 11 Countries and Took Taipei Emergency Dispatch Records

The Story: 

Acronis's Threat Research Unit says one attacker, likely based in Hualien City, Taiwan, ran a setup of AI agents called BOSSAI against hundreds of organizations in at least 11 countries. The agents collected data or accessed systems at about 10 organizations in Taiwan, Mexico and Thailand. The data included case records from Taipei's "119" fire and ambulance dispatch system for seven districts. The Hacker News reported the findings on November 9.

The details:

  • BOSSAI had 12 agents, 9 of them in use. Some built and maintained servers and tools, while others scanned and attacked targets. The agents did not talk to each other, so the attacker pasted output from one into another by hand.

  • Commercial cloud models did most of the building, with Grok as the main builder, and Grok and MiniMax M3 refused direct attack requests. The attacking agents ran mostly on "abliterated" local versions of DeepSeek, GLM and Qwen with safeguards removed. The attacker also redirected Claude Code and Codex to such a model, mostly for infrastructure work.

  • The dispatch records, pulled through the system's API, included incident times, response types, case status and street-level locations. The agents also found a Taipei fire department platform exposing 2018 patient data, including ECG readings, with no login. Elsewhere, three Thai universities were breached through SQL injection and about 206,800 Mexican home phone records were assembled.

  • The tooling was far from perfect. One scanner invented CVEs, more than half of its high and critical findings were false positives, and the main agent reported results that never happened. Researchers watched the operation for 12 days through a server holding more than 90,000 files.

Why it matters: 

Safety refusals on commercial models did not stop this attacker; he moved the offensive steps to local models with safeguards stripped out. Model-side guardrails are one layer, not a control enterprises can count on. Read with the Operation Snowball story, the pattern is clear: agents cut the labor of an attack even when they make mistakes. Defenders should expect noisier, higher-volume campaigns against exposed APIs and default credentials, and should check which model endpoints their own coding agents are allowed to call.

🔗 Fireworks AI Breach Spreads to Cursor, Notion and a Trojanized Claude Code Extension

The Story: 

Cursor said an attacker used an access key it had issued to Fireworks AI, exposed in the Fireworks breach, to copy data from a storage bucket the two companies shared. With credentials found in that data, the attacker placed trojanized builds of Anthropic's Claude Code extension in the cache serving the Cursor Marketplace on October 29. The builds were designed to install a persistent SSH backdoor and stayed available for about 11 hours. The Hacker News reported Cursor's update on November 15, five days after Fireworks confirmed its own breach was wider than first disclosed.

The details:

  • Fireworks detected unauthorized use of internal credentials on October 6. Its later update says the attacker used a third-party cloud provider hosting R&D machines to reach its development environment on October 23 and ran a workload on October 24 that affected customer data. The exposed data includes billing details, secrets, and some incomplete model weights and datasets.

  • Fireworks attributes the activity to a financially motivated group that sells LLM access from stolen credentials to illicit routing gateways, a practice known as LLMjacking. Notion said fewer than 1% of workspaces were affected through its Insights feature, which processed Notion AI chat transcripts. Notion disabled Insights and stopped serving open-weight models through Fireworks.

  • Cursor says at most 5,000 users who requested specific cached versions on October 29 may have received the malicious build, versions 2.1.312 or 2.1.313. The shared bucket held coding agent traces from sessions not in Privacy Mode, which exposed API credentials for less than 0.01% of customers, plus archived non-production repositories and support data.

  • Cursor moved all inference traffic off Fireworks, revoked the vendor key and took its marketplace offline during cleanup. It says no further malicious activity has been seen since November 1. Uber said in a California filing that a service provider's compromised credentials gave access to its data on October 9 and 13; TechCrunch reported the provider was Fireworks, which Uber has not confirmed.

Why it matters: 

One inference provider's stolen credentials turned into data loss at several AI companies and malware on developer machines. Most enterprises do not list their model hosting providers as critical third parties, yet those providers see prompts, agent traces, secrets and fine-tuning data. The attackers' goal was reselling LLM access, so stolen keys are a direct cost line as well as a breach risk. Security teams should inventory every inference provider, rotate keys shared with them, and make sure agent traces never store plaintext credentials.

☁️ Microsoft 365 Outage Takes Down Outlook, Teams and Copilot in More Than 70 Countries

The Story: 

A Microsoft 365 outage that began on the afternoon of November 12 (UTC) cut access to Outlook and Exchange Online worldwide and partly disrupted Teams and Microsoft 365 Copilot. More than 12 hours in, Microsoft said it had only "marginal success" restoring service. Downdetector logged at least 43,000 Outlook reports from 72 countries by Friday morning. Microsoft said there was no indication of malicious or cyber activity.

The details:

  • Microsoft acknowledged mailbox access problems in the Americas at 17:54 UTC on November 12 and warned of global impact minutes later, under incident MO1550821. It reported partial recovery at 04:36 UTC on November 13. Wikipedia, citing the BBC, says services were fully restored on the morning of November 14.

  • Microsoft described servers failing automated health checks and dropping out, which overloaded the servers that remained. It said the problem may come from a component that processes server health, and it rolled back a recent network configuration change without saying whether that change was the cause. No post-incident review had been published as of November 13.

  • Affected organizations included Japan's Government Solution Service, the Supreme Court of the Philippines, Singapore courts, Toronto hospitals and parts of the UK NHS, where some staff went back to paper records. Al Jazeera reported that integrations used by OpenAI and Anthropic were also affected.

  • Anderson Economic Group estimated about $7 billion in lost productivity, with roughly 430 million people using Microsoft 365. The incident follows a nearly 10-hour North American outage in January and a September incident that Microsoft blamed on a core authentication configuration.

Why it matters: 

Copilot and agent workflows run on the same identity, mail and collaboration layer that went down, so the outage stopped AI automation along with email. Availability of AI agents is a business continuity question, not only a model question. Agents that act on email or Teams data need defined behavior when their tools fail: pause, queue or fail closed. Read with the Fireworks story, dependence on a few AI and cloud providers is now a board-level risk.

🎙️ Anthropic Launches Claude Meetings, Putting an AI Agent Inside Live Video Calls

The Story: 

Anthropic launched Claude Meetings in beta on November 10, letting Claude join Google Meet calls as a visible participant on Team and Enterprise plans. During a call, Claude can draft the plan being discussed, map a timeline on a whiteboard, answer chat questions and pull information from connected work apps. Notes, transcripts and action items are saved to a meeting page in Claude afterward. Anthropic launched Claude Whiteboards and Claude Sheets in beta the same day.

The details:

  • Claude records a transcript, a summary and descriptions of shared screens, but does not save audio. Google Meet is the only supported platform at launch, and support for other video platforms is planned.

  • Claude announces itself and posts a notice in the meeting chat. By default it will not join calls with external guests and leaves if one joins; an admin can allow it to stay, in which case it does not reply in chat but guest comments still appear in the notes and transcript.

  • On Team plans, inviting Claude is on by default and admins can turn it off; on Enterprise plans, an admin must turn it on. Each user approves Claude's use of their connected tools the first time. The beta excludes customers using zero data retention, HIPAA-ready configurations or customer-managed encryption keys.

  • Notes and transcripts carry no extra charge during the beta, and transcripts can be sent to tools such as Gong, Granola, Fireflies.ai and Otter. Anthropic advises users to make sure everyone on a call is comfortable being transcribed before adding Claude.

Why it matters: 

Meeting assistants are turning into agents with access to company apps, not just note-takers. That puts confidential conversation, screen contents and connected data in one place, which makes meeting policy part of AI governance. The defaults here (leaving when external guests join, per-user tool approval, admin opt-in on Enterprise) are a useful baseline for judging any meeting agent. Security teams should decide who can invite agents to calls, where transcripts flow next, and how long they are kept.

💸 Google's Gemini 4 Hydrogen Runs 2.5x Faster Than Gemini 3.8 Flash at Double the Price

The Story: 

Google released Gemini 4 Hydrogen on November 12 across the Gemini API, AI Studio, the Gemini app, Google Search and Gemini for business. Google says it completes tasks 2.5 times as fast as Gemini 3.8 Flash at its default reasoning setting. Standard pricing is $1.50 per million input tokens and $7.50 per million output tokens, double Flash's current introductory rate. The launch landed in the same week Anthropic released Claude Fable 5.2 at a 20% lower price.

The details:

  • Hydrogen scores 51 on the Artificial Analysis Intelligence Index at high reasoning, versus 31 for Gemini 3.8 Flash. On Google's Terminal-Bench 4.1 it completes 57% of tasks, versus 18.2% for Flash and 63.9% for Claude Sonnet 5.5 in Google's own table.

  • Cached input costs $0.075 per million tokens, and the Batch, Flex and Contributor tiers cost half the standard rate. Contributor is invite-only and lets Google use request data for product and model training. VentureBeat notes Hydrogen costs 15 times as much per token as the cheapest small models from Anthropic, OpenAI and Meta.

  • Default reasoning is set to high, with a new xhigh level, and the context window of about 1 million tokens is shared by input, output and reasoning. In Box's early-access test, accuracy on its Complex Work Eval rose from 74% to 79%, with tasks taking under half the time and about 12% fewer tokens.

  • Anthropic priced Claude Fable 5.2, released November 9, at $8 and $40 per million tokens, down from $10 and $50, and it completed 76.3% of Irregular's multi-stage CyScenarioBench challenges versus 61% for GPT-6 Cyber. Pricing trackers also report that Gemini 4 Carbon and Argon list prices rise from $2/$10 to $4/$20 on December 1.

Why it matters: 

Faster, stronger mid-tier models are arriving at higher unit prices, and a higher default reasoning level quietly raises token use per task. AI cost control now depends on routing each task to the cheapest model that can do it and watching reasoning settings, not just list prices. Data-sharing tiers like Contributor trade a lower price for training rights, which security and legal teams should review before anyone opts in. With several price changes landing on December 1, finance and platform teams should re-check AI budgets this month.

What’s next?

Thanks for reading! If this brought you value, share it with a colleague or post it to your feed. For more curated insight into the world of AI and security, stay connected.