Opening

Blockchain analytics firm Chainalysis has publicly admitted that external attackers used "offensive agentic techniques" to gain unauthorized access to internal systems, leaking customer names and email addresses, and the company is explicitly warning everyone to watch out for follow-up phishing. Today I want to walk you through three things: exactly what data this AI-driven intrusion touched, a Brotli remote code execution bug called HTTP2Shell that an AI agent dug up in a single afternoon, and why the AI coding assistants on your dev machines have become an attack surface your scanners can't see. I'll tell you how to inventory it shortly.

Today's AI Security Highlights

1. Chainalysis Breached via "Offensive Agentic" Techniques, Basic Customer Data Leaked

• What happened: On October 30, blockchain analytics firm Chainalysis disclosed that external attackers used "offensive agentic techniques", meaning AI agents driving the attack workflow autonomously, to gain unauthorized access to internal systems and tools. The accessed data included non-sensitive customer information such as user IDs, names, email addresses, and organizational role labels. According to Infosecurity Magazine, the company said it found no evidence of plaintext passwords, MFA secrets, or session tokens being exposed, and clarified that the incident is unrelated to the Binance event. The vulnerability has been patched and systems hardened.
• Scope of impact: Affected parties are Chainalysis platform account holders. Names, emails, plus organizational role labels are the perfect combination for targeted phishing. When attackers know who the admins are and who owns what, fake password resets and MFA prompts become significantly more effective.
• Defensive recommendations: If your company is a customer, follow the official guidance and enable two-factor authentication. Add this to this week's reminders: never approve password reset emails or MFA prompts you didn't initiate yourself, and confirm through internal channels instead. At the same time, add a monitoring rule on your email gateway for lookalike sender domains impersonating the brand.
• Source: Infosecurity Magazine (https://www.infosecurity-magazine.com/news/chainalysis-customers-offensive/)

2. HTTP2Shell: An AI-Discovered Brotli RCE Whose Patch Sat Upstream for Nearly Eight Months

• What happened: HTTP2Shell (CVE-2026-111497) is a remote code execution vulnerability in the Brotli compression library, discovered by an OpenAI agent during a high-compute red team exercise in late September. The awkward part: the underlying fix actually landed in the Brotli source code back on February 27, but because it wasn't recognized as a security patch, it went nearly eight months without a release or advisory. Help Net Security published an analysis on November 3.
• Scope of impact: Brotli is pulled in by widely used components including Node.js, Bun, Envoy, and Python, which means everything from your frontend gateway to your backend services may depend on it indirectly. The real risk here is visibility: the code was fixed but no advisory was published, so during that window your SCA and vulnerability scanning tools would never have told you that you were running an affected version.
• Defensive recommendations: Start by using your SBOM or dependency manifest to pull out every place Brotli is used and which version, then schedule upgrades. Until a proper patch is in place, you can follow the article's approach and add runtime rules that block behavior that shouldn't occur, such as a compression handling flow spawning a new process, as a virtual patching mitigation.
• Source: Help Net Security (https://www.helpnetsecurity.com/2026/11/03/product-showcase-sweet-security-virtual-patching/)

3. AI Coding Assistants as a New Attack Surface: Copilot RCE Patched, Two Other Cases Not Treated as Vulnerabilities

• What happened: Security research firm Adversa AI's November roundup on AI coding agents notes that a single trailing dot in a filename was enough to make Copilot inside Visual Studio and SSMS write its own MCP settings and configuration files, leading to command execution. Microsoft rated CVE-2026-103572 as Important and has patched it. MCP is the interface protocol AI agents use to connect to external tools, so rewriting the config file effectively means rewriting tool permissions. The article does not carry a publication date.
• Scope of impact: The other two reports ended differently. Anthropic classified a report about sub-sessions bypassing hook consent as Informative, and GitHub confirmed the Copilot CLI finding but did not treat it as a vulnerability, shipping the fix only in 1.18.33 with no CVE and no advisory. No CVE means scanners won't flag older versions, so the risk on developer machines falls entirely on your own inventory capabilities.
• Defensive recommendations: Add AI coding assistants on developer laptops to your asset inventory, confirm that Visual Studio and SSMS have Microsoft's patch applied and that Copilot CLI is upgraded to 1.18.33 or later, and periodically review agent MCP configuration files for unexpected modifications.
• Source: Adversa AI (https://adversa.ai/blog/top-ai-coding-agent-security-resources-november-2026/)

4. Security Budgets Are Up but Nobody Owns Shadow AI: 46% Believe Unapproved AI Is Connected to Company Systems

• What happened: Security vendor DNSFilter released a report on November 3 noting that while enterprises have increased security spending, they haven't closed the shadow AI governance gap. In the survey, 46% of IT and security staff believed their company systems were already connected to unapproved AI applications, and one in five respondents said unapproved AI integrations had already caused a security incident. Cybersecurity Dive covered the report on November 4.
• Scope of impact: This one has nothing to do with company size. As long as employees can install browser extensions on their own, wire up SaaS integrations, or plug an API key into a development environment, you'll have invisible data exfiltration paths. The protection you paid for can't cover connections you don't know exist.
• Defensive recommendations: The order matters: inventory first, then control. Pull the AI applications and integrations actually in use from your network egress logs and SaaS OAuth authorization lists, then set policy and access controls. Don't do it backwards by issuing a ban first.
• Source: Cybersecurity Dive (https://cybersecuritydive.com/news/ai-security-governance-spending-dnsfilter/873725)

Closing

The one thing worth remembering today is that no CVE and no advisory does not mean no risk. The Brotli fix sat unnoticed for eight months, and the Copilot CLI fix never got an identifier. In both cases your scanner would never have warned you. Do one thing today: list out your AI-related assets, including coding assistants on dev machines, SaaS AI integrations, and OAuth authorizations. See it first, then talk about controlling it. See you tomorrow.