Anthropic's stage-by-stage playbook for the AI-native SDLC: how teams plan, design, build, test, deploy, and maintain software with Claude.

GPT-5.6-Cyber, given a QEMU/KVM VM sandbox, escaped three times: exploiting a known kernel CVE, chaining two libslirp vulnerabilities, and finally using four bugs including three 0-days. Standard VMs are insufficient to contain cyber-capable AI agents; Firecracker with minimal attack surface is recommended.

Reactive bug-patching (whack-a-mole) is a losing security strategy. Instead of endlessly finding and fixing individual vulnerabilities, security teams should define and enforce system invariants to eliminate entire bug classes, including in the AI-accelerated vulnpocalypse era.

Practical LLM evaluation lessons from GitHub secret scanning: define product goals and guardrails first, treat offline evaluation as repeatable integration testing, keep eval data close to production, audit labels, use error analysis, and apply LLM-as-judge for human review triage.

A stored XSS flaw (CVE-2026-13323) in Open VSX allowed an attacker to mint a persistent PAT in a maintainer's session via one click, then silently publish malicious extension updates to millions of developer endpoints via auto-update.

August 2026 census of the top 1M domains: DMARC valid 63.6%, DANE 4.6%, MTA-STS 1.2%, BIMI 2.5%. New finding: only 0.2% of MX servers actually refuse cleartext, despite published policies. DMARC leads at 63.6%.

A browsable directory of 83 AI data centers tracked by Epoch AI, covering hyperscale campuses, colocation sites, and AI training facilities across North America, Europe, and Asia. Each entry includes IT power capacity, compute estimates, operator details, satellite imagery, and buildout timelines.

Malicious versions of the arrayref Rust crate (and others) executed a backdoor at compile time. The campaign's infrastructure overlaps with recent DPRK supply chain attacks, including Mastra and axios.

Like netcat, but over Tailscale's data plane, without Tailscale's control plane.

Boost wraps the commands your agents already run, turning noisy logs into compact, structured context that keeps the signal while cutting the noise.

A guide to detecting multi-stage AWS attacks by correlating signals across CloudTrail, VPC Flow Logs, and Route 53 DNS logs. Covers four business-context-aware patterns: unexpected S3 access, abnormal role chains, KMS key misuse, and off-hours privileged changes, with CloudWatch Logs Insights queries and Lambda automation.

A guide for migrating ClickOps-managed AWS infrastructure to governed CloudFormation IaC using IaC Generator for template generation, stack organization by lifecycle/ownership, and automated drift detection via EventBridge for continuous compliance monitoring.