AI, explained properly
Deep dives, comparisons and plain-language explainers on artificial intelligence — for people who want substance over hype.
PHP2Shell CVE-2026-109618 and CVE-2026-109619 — CVSS 9.2 — critical PHP RCE leaked before patches. Unauthenticated remote execution via multipart uploads; proof-of-concept already circulating in the wild.
Anthropic's DevDay release pairs Claude Sonnet 5.5 ($2/$10) with Claude Code Mods — TypeScript customization hooks for agentic workflows — directly responding to OpenAI's pricing and agent architecture with a customization-first strategy.
The FTC expanded its AI investigation to include independent evaluators METR and Redwood Research. What it means for agent oversight, independent evaluation, and regulatory escalation.
OpenAI Dots represents a shift from chat-only to persistent autonomous agents — each Dot gets its own cloud computer and browser, working continuously after the user walks away, with multi-agent orchestration across ChatGPT and Codex.
Anthropic's Claude Code v2.1.289 (Oct 3) fixes a real permission bypass where sh -c 'rm -rf X' bypassed Claude Code's deny/ask checks. Six security updates — bash enforcement, symlink restrictions, MCP sign-in — critical for agentic coding users.
California Attorney General Rob Bonta issued the first state-level investigative subpoena targeting AI model security failures on Sept 30, following the July Hugging Face breach. The legal escalation signals that rogue agent incidents are moving from internal safety matters to state liability — full details for OpenAI Safety.
OpenAI's longest-tenured safety leader, David Robinson, resigned and published an essay in The Atlantic arguing that OpenAI's 'iterative deployment' culture — find problems, patch, release — guarantees periodic failures as models grow more capable, and that the failures are cultural rather than procedural.
OpenAI's DevDay 2026 release of GPT-6.1 Sol delivers frontier-level agentic coding and professional work at one-fifth the price of GPT-6 Astra — $2 input, $10 output per million tokens — with a 1.05M-token context window and multi-agent beta support.
GPT-6 Astra pushes multimodal reasoning to the edge with 87.5% on video QA, 91.2% on audio benchmarks, and a 60B-parameter Navier edge stack running on laptop GPUs.
September 2026 safety stack: NVIDIA's OpenShell + Sentry on BlueField-4, OpenAI's GPT-6.1 Astra cancellation, Anthropic's threat report showing 80-90% autonomous agent attacks.
DevsIsle designs and ships AI systems, agents and integrations for teams that need it done properly.
