AI Security - Netics

Sign in Subscribe

AI Security

A collection of 12 posts

Netics editorial card for the analysis of Anthropic's open-source vulnerability disclosure ledger.

Anthropic's Open-Source Disclosure Ledger Shows Where Remediation Slows Down

Anthropic's dashboard now reports 6,157 disclosed open-source vulnerabilities against 516 patches released upstream. The funnel behind those two numbers, and the pacing policy the vendor app

Netics editorial card on the Mythos vulnerability programme, showing the disclosure queue and the shortened exploitation window.

Claude Mythos Found a Critical HFS Bug, and Exploitation Followed Within a Day

Anthropic's vulnerability programme has disclosed 6,157 findings across 591 open-source projects. The HFS session-forgery bug shows what the programme's throughput means in practice: a publi

Netics editorial card for the Anthropic September 2026 threat-intelligence analysis, with the official Anthropic identity.

Stolen AI API Keys Are Now the Attack Currency

Anthropic's September 2026 threat report documents a criminal AI supply chain built on stolen API keys and session tokens. One hacktivist campaign ran for a month entirely on stolen keys; th

Netics editorial card for the Perplexity SPACE agent-sandbox red-team analysis, with the official Perplexity identity.

Perplexity Red-Teamed Its Agent Sandbox: The VM Held, the Network Didn't

Perplexity's Secure Intelligence Institute published a red-team report on its own SPACE sandbox: zero VM-to-host escapes in 108 runs, but four models bypassed the network egress policy, and

Netics editorial card: the network enforcement point between users, agents, and AI apps.

Microsoft's Network DLP Now Watches AI Agent Uploads. Roll It Out in Audit Mode

Purview network DLP on Entra Global Secure Access blocks sensitive data at the network layer — including traffic an AI agent sends on a user's behalf. The licensing and limits deserve scruti

Netics feature card for the Meta Muse Secure VM article using official Meta Newsroom imagery

Meta's Muse Tries to Make the Consumer Agent a Secure Computer

Meta's Muse personal agent ships with a dedicated Secure VM and a separate Sentinel approval agent. Netics explains what the architecture gets right and what it still leaves open.

Netics editorial feature card showing interacting software agents, shared context, and control points for enterprise multi-agent security.

Microsoft Red-Teaming Shows What Breaks When Agent Networks Interact

TL;DR Microsoft Research tested more than 100 autonomous agents in a shared platform where they could post publicly, send direct messages, use applications, schedule meetings, exchange currency, and trade goods. The research found that a reliable individual agent is not a reliable network: messages can propagate, reputation systems can

Netics security visual showing a normal funding lure word split by invisible U+E0020 Unicode tag space before detector analysis.

Invisible Unicode turned an AI evasion trick into a phishing filter test

Microsoft’s February 2026 telemetry shows invisible Unicode can fracture phishing keywords. Netics explains normalization, blind spots, layered controls, and telemetry limits.

Illustration of AWS AI security layers and deployment phases.

What the AWS AI Security Framework Actually Verifies

AWS's new AI Security Framework maps controls across three use cases, three layers, and three phases — a genuinely useful map, but not proof that any given workload is actually secure.

Google’s ADK zero-trust example shows why agent safety needs signed writes, sandboxed code execution, and deterministic gateways outside the model.

Zero-Trust AI Agents Need Hard Boundaries Outside the Model

Google’s ADK zero-trust example shows why agent safety needs signed writes, sandboxed code execution, and deterministic gateways outside the model.

NVIDIA AI agent security thumbnail with a licensed AI-compute hardware photo background and readable Netics title.

In an AI Agent Stack, the Runtime Must Own Security

Prompts and harnesses can guide an agent, but authoritative identity, policy, isolation, effect checks, and audit belong below the modifiable agent boundary.

Cloudflare AI gateway article card with the official Cloudflare symbol and Cloudflare name on a dark Netics card

Don't buy an AI gateway before you know what your agents touch

MCP sprawl is the real AI governance risk, but the gateway-first answer is a vendor's, not yours. Inventory what your agents can touch before buying anything.