We, as an industry, should have the courage to throw away the playbooks we've been working on for the past 20+ years and create new (and better) ones for the AI era.

GitLab's CISO argues that agentic AI development requires security, governance, and guardrails embedded in the SDLC execution path. The key metric is time from detection to verified remediation, driven to machine speed using AI-powered triage, constrained agent identities, and continuous scanning.

Test-driving Jev, a decision model trained for calibrated probabilities, on identity resolution across five data sets. Compared against gpt-oss-20b, Claude Haiku 4.5, Sonnet 5 and Opus 5.5, and traditional entity resolution (Fellegi-Sunter and gradient-boosted trees), with cost, latency, prompt, stability and error analysis.

BeyondTrust found two RCE CVEs in the Amazon Bedrock AgentCore Python SDK's install_packages() helper. A crafted package name with a newline or pip extras command substitution bypassed input validation, achieving RCE inside the Firecracker sandbox and stealing attached execution role AWS credentials from MMDS.

Cloudflare Application Profiles launches a positive security layer for web apps: it learns HTTP request structure from traffic, flags deviations (type mismatches, out-of-range values, invalid UUIDs), and exposes violations as metadata for Security Rules enforcement.

DigitalOcean Managed Agents secure autonomous workloads by ensuring agents never hold live credentials directly, utilizing ephemeral execution-time brokering via Action Gateway and strict runtime boundaries. This architecture aligns with NVIDIA's Open Agent Safety Platform and OpenShell policy schemas to prevent data exfiltration and limit blast radii across agent swarms.

By combining agentic reasoning with live-site verification, PageBreak uncovered complex logic flaws that traditional scanners will often miss. The post shows three examples of high-severity vulnerabilities PageBreak found in Google applications.

Datadog Security Labs found GHSA-632h-h47v-g4x4 in OpenCode: the unauthenticated /global/upgrade API endpoint accepted arbitrary npm tarball URLs, enabling RCE via a crafted cross-origin HTML form submission bypassing CORS.

A GitHub CLI extension to enable security features on repositories following best practices from GitHub Security Lab.

Isolated VM environment for running Claude Code and Codex.

The agentic CLI for the entire Cloudflare API.

A simple touchID prompt'er for use in shell scripts.

Storm-3168 (JADEPUFFER) used compromised Azure service principals to conduct automated reconnaissance, bulk deletion of Storage Accounts, Key Vaults, and Function Apps, and credential harvesting via ListKeys, consistent with ransomware-aligned objectives. Mitigations include least privilege, secret rotation, and resource locks.