ARTEX Accelerates Cyberattacks; Snyk Joins OpenAI's Defense Factory
ARTEX Accelerates Cyberattacks; Snyk Joins OpenAI's Defense Factory
AI Sec News Weekly #35 — 333 sources scanned
AI is accelerating both sides of cybersecurity. ARTEX-linked intrusions reached privileged access in under three hours, while OpenAI's new Defense Factory brings Snyk security findings into agent-driven workflows for vulnerability investigation and remediation. The race is no longer just about finding vulnerabilities faster, but fixing them before attackers exploit them.
Elsewhere, agent-security testing gets more concrete, AI-generated proofs connect protocol claims to implementation, and a browser-agent developer confronts the limits of borrowing a user's identity.
This Week's Stories
FBI Links ARTEX Agents to Global Attacks and ConnectWise Breach
The FBI says threat actors have used ARTEX since at least August 2026, including in a ConnectWise supply-chain breach affecting Microsoft 365 customer environments. The Go-based framework uses external LLMs to drive reconnaissance, exploitation and data theft. Some intrusions reached privileged access in under three hours; researchers counted over 1,600 exposed ARTEX consoles in early November.
Why it matters: Detection alone may offer little protection when an attacker can finish escalating privileges before defenders respond.
BleepingComputer by Lawrence Abrams — Published 2026-11-15
Snyk Joins OpenAI's Daybreak Defense Factory
Snyk is among the first integration partners for OpenAI's Daybreak Defense Factory, an agent-driven approach to continuously finding, validating and fixing vulnerabilities. Snyk findings feed into the discovery stage, bringing dependency paths, reachability analysis and remediation context to AI agents investigating potential weaknesses. Proposed fixes return through existing pull request and ticket workflows, with human approval and independent verification remaining central to the process.
Why it matters: Finding vulnerabilities faster only helps if teams can determine which ones are exploitable and safely fix them. Combining security-tool context with agent-driven remediation could help close that gap.
Snyk Documentation — November 2026]
Tool Spotlight
New tools and approaches for AI security work.
ai-agent-security-toolkit Puts Prompt Injection and Tool Misuse Under Test
The Python package ai-agent-security-toolkit combines 500 canary-scored injection cases, a deny-by-default tool guard, and output validation. Its taint tracking blocks side-effecting tools after untrusted content is read—a useful test case for document assistants with file-write access. CI can fail on attack-success rates; the published deterministic simulator results are not an LLM benchmark.
Why it matters: Prompt-injection resistance becomes a testable property of the surrounding application, rather than a bet on the model refusing malicious instructions.
pypi.org — Published 2026-11-15
AlphaProof Code Connects Signal Security Claims to Checked Lean Proofs
DeepMind's AlphaProof Code uses human-guided agents to write Lean specifications and proofs checked by the Lean kernel. Applied to Signal's PQXDH handshake and Triple Ratchet, it generated 6.5 million lines of Lean to connect protocol guarantees to libsignal's implementation. The result assumes perfect cryptography; it does not establish computational secrecy of bytes.
Why it matters: Formal-methods engineers still own the hardest judgment: whether the specification captures the security property they actually care about.
The Register Security by Thomas Claburn — Published 2026-11-10
Community Chatter
What practitioners are debating.
Token’s Browser Agent Complicates Its Own Agent Identity Advice
Token says its Rehearsal browser agent broke its own security rules by default: it inherited the user’s session rather than having a separate identity, limited access and independent traceability. Token added action recording, read-only dashboards and approval before saves. The catch, its authors acknowledge, is that they control both the agent and the application—an advantage outside browser agents don’t share.
Why it matters: A valid session cookie establishes account access, not the user’s consent to each action an agent takes.
The Hacker News — Published 2026-11-11
Schneier’s AI Hacking Argument Goes Beyond Bugs and Jailbreaks
In an essay adapted from his August 7 DEF CON 34 keynote, Bruce Schneier argues that autonomous AI could exploit tax rules, financial regulations and political systems without malicious intent. His definition of hacking includes behavior a system permits but its designers never wanted. He ranks AI exploiting other systems above attacks on AI itself—a broader framing than software security.
Why it matters: Security evaluations centered on jailbreaks would miss the highest-risk behavior in this threat model.
Schneier on Security by Bruce Schneier — Published 2026-11-13
Quick Hits
- Cursor Breach Led to Malicious Claude Code Extensions (BleepingComputer; Published 2026-11-14) — Cursor says attackers used credentials exposed in coding-agent traces to publish malicious Claude Code extensions that attempted to install persistent SSH backdoors.
- Microsoft Backdoors AI Agents by Editing Fewer Than Ten Neurons (The Hacker News; Published 2026-11-11) — Researchers edited fewer than 10 neurons in a model checkpoint to make AI agents run attacker-chosen commands.
- Claude Code Settings Files Exposed Thousands of GitHub Secrets (Dark Reading; Published 2026-11-10) — Apiiro found 2,539 credentials in roughly 2,000 public GitHub repos, tracing leaks to secrets Claude Code saved in
.claude/settings.local.json. - Project Zero Plans Shorter Disclosure Deadlines as LLMs Accelerate Attacks (Google Project Zero; Published 2026-11-13) — Citing faster LLM-assisted attacks, Project Zero plans to cut deadlines from 90 days to 60 in February 2027 and 30 in January 2028.
- Three Arrested in Alleged AI Impersonation Fraud Scheme (Schneier on Security; Published 2026-11-11) — Three people were arrested in an alleged $129 million-plus fraud using AI-generated voices and images to impersonate executives and lawyers.
- IBM Releases Granite Previews With Prompt Injection Defense Format (Hugging Face Blog; Published 2026-11-12) — Granite 4.1 previews include LLMON, a format separating instructions from data that reduced prompt-injection success in IBM’s tests.
- llama.cpp 1.0 Gives Local Inference Libraries Explicit Compatibility Rules (Hugging Face Blog, Daniel Bevenius; Published 2026-11-13) — llama.cpp and ggml v1.0.0 promise major-version bumps for public API and ABI breaks: clearer upgrade signals for offline assistants handling sensitive incident reports, not a security fix.