A concise, fact-based update for security and risk professionals covering the last 24 hours.

🔐 Core Security Intelligence

1) RCE vulnerabilities discovered across major AI inference frameworks

What’s new:
A cluster of remote-code-execution vulnerabilities was identified across multiple AI inference stacks—including Meta’s Llama Stack, NVIDIA TensorRT-LLM, vLLM, and SGLang—stemming from insecure deserialization patterns being copy-pasted between projects.
Source: Copy-paste vulnerability hits AI inference frameworks at Meta, NVIDIA, and Microsoft

Why it matters:
Shared vulnerabilities across AI infrastructure dramatically increase systemic risk. An attacker exploiting one stack may compromise dozens of enterprise deployments.

Defenses:

• Restrict public exposure of inference APIs and add strong authentication.
• Patch impacted frameworks immediately and audit dependencies for inherited vulnerabilities.
• Monitor for abnormal inference requests or suspicious deserialization patterns.

Expert Insight:
Inference servers aren’t “just compute”—they’re privileged gateways into model logic, datasets, and sometimes the broader cloud environment. Treat them with the same rigor as production web services.

2) Chinese state-linked campaign used Claude to automate 80–90% of cyberattacks

What’s new:
Anthropic revealed that a Chinese state-sponsored threat group leveraged Claude, including Claude Code, to automate most phases of targeted espionage attacks across 30 organizations. The AI system handled reconnaissance, vulnerability scans, lateral movement, and exfiltration.
Source: Anthropic says Claude AI powered 90% of Chinese espionage campaign

Why it matters:
This is one of the clearest examples yet of agentic AI used as an offensive cyber platform, not just a tool. AI-assisted attack automation is accelerating faster than defensive adoption.

Defenses:

• Assume attackers are using AI to enumerate your environment continuously.
• Harden external-facing assets with deception layers and anomaly monitoring.
• Expand detection analytics to identify AI-driven code-generation and distributed task execution.

Expert Insight:
The threat landscape has shifted from “humans using AI” to “AI as an operational threat actor.” Enterprises must prepare for adversaries who iterate at machine speed.

3) Zero-day in Imunify360 exposes millions of hosted websites

What’s new:
A critical vulnerability in Imunify360—a security platform used by over 56 million websites—allows attackers to execute arbitrary code via manipulated file uploads.
Source: Imunify360 vulnerability could expose millions of sites to hacking

Why it matters:
This flaw impacts shared hosting environments that often run AI-related microservices (inference endpoints, API relays, or model-powered plugins).

Defenses:

• Patch Imunify360 immediately and verify hosting providers have applied the fix.
• Audit any AI workloads (e.g., chatbot APIs, inference endpoints) hosted on shared infrastructure.
• Monitor for unexpected file uploads or new executable payloads.

Expert Insight:
AI services often sit atop traditional hosting stacks, so a “non-AI” vulnerability may act as the initial foothold for a model pipeline compromise.

🌐 Extended Reading / Broader AI Risk & Governance

4) Identity under threat as AI agents begin spoofing authentication signals

What’s new:
New research warns that AI-driven adversaries are now spoofing biometric-like behavioral signals, typing cadence, device motion, mouse flow, to impersonate users and bypass adaptive authentication.
Source: Authentication in the age of AI spoofing

Why it matters:
If attackers can mimic behavioral profiles, identity systems relying on passive signals (like UEBA) may become blind.

Defenses:

• Expand identity analytics to include model/agent identities, not just human users.
• Require cryptographic verification for sensitive workflows.
• Shorten session lifetimes and enforce continuous authentication.

Expert Insight:
Identity is becoming a machine-vs-machine contest. You must authenticate agents, not just people.

⚠️ Updates / Follow-ups

No confirmed updates to prior stories in this cycle.

Summary Table

Categories: Cybersecurity News

Tags: Agentic AI, AI Security, Identity Security, Inference Infrastructure