Anthropic's Deputy CISO shares a four-question framework for assessing agentic AI risk, and walks through controls that keep agent deployments bounded and auditable.

An autonomous AI agent exploited two RCE paths in HuggingFace's dataset processing pipeline, escalated to cluster level, harvested credentials, and moved laterally. Defense forensics used self-hosted GLM 5.2 after commercial APIs were blocked by safety guardrails. You can also read OpenAI's post on this.

A guide to open-source AI tools for finding code vulnerabilities, comparing exploit generation, skill-boosted auditing, and SAST+LLM hybrid approaches.

A reference architecture for giving a locally-running AI agent a trustworthy, auditable identity, without long-lived credentials on disk, including a structural defense against prompt injection built into the protocol layer.

Two chained bugs in Google's OAuth device code flow: transferable sign-in sessions via URL, and missing server-side client_id/scope binding to device_code. Combined with prompt=none, any victim opening one link silently yields tokens for arbitrary Google-registered clients, including Gmail.

Attackers compromised a PHP developer's GitHub repos, injecting 583 malicious Actions workflows across 10 Packagist packages. These workflows used GitHub-hosted runners to exploit CVE-2026-41940 (cPanel/WHM auth bypass) and exfiltrate server credentials.

A study of ~200,000 LLM responses found 5 frontier models (Claude, GPT, Gemini, DeepSeek) hallucinate nonexistent package names at 4.62-6.10%, with 53 shared fictitious names on PyPI/npm still registrable and exploitable via slopsquatting attacks.

Cloudshell boto3 hidden API.

Zeedumper connects to a Kubernetes cluster using your kubeconfig and dumps component z-pages (flagz, statusz, and configz) retrieving them through the API server proxy. You can also check out the companion blog post.

An egress firewall for untrusted workloads.

Amazon GuardDuty investigation agent (public preview) uses AI to auto-investigate GuardDuty security findings, reducing investigation time from hours to minutes. It returns risk levels, confidence scores, MITRE ATT&CK mappings, and remediation steps via console, CLI, API, or AWS MCP server.

Amazon announced the Claude apps gateway for AWS, a self-hosted control plane that gives organizations a single point of control over access, cost, and policy for Claude Code and Claude Desktop.