Blog
Writing about AI — how it works, where it helps, and what we learn building with it.
Latest posts
16,326 Supabase Databases Are Readable by Anyone. The Common Thread Is a Table an Agent CreatedUpGuard found 16,326 Supabase-backed databases whose tables can be read by anyone on the open internet, over half of them holding personal data. Nothing was breached — the exposure is a default that applies on the path coding agents take, and not the one humans take.Security
Kiteworks Told Customers to Switch Their Servers Off Tonight. There Is No CVE, No Patch, and No Confirmed BreachOn September 25 Kiteworks emailed customers telling them to power down their file transfer servers for a six-hour window, on the strength of threat intelligence passed to it by federal authorities. There is no CVE, no confirmed compromise, and no patch — and the advice covers servers that are not reachable from the internet at all.Security
Bitget Lost $351.6 Million Without Losing a Private Key. That Is the Detail Worth Reading TwiceRoughly $351.6 million left Bitget's hot and warm wallets across seven blockchains on September 24. The exchange says no private key was stolen — attackers compromised the backend that tells the signers what to sign, which is a boundary the industry's hot/cold custody model does not defend.Security
Cloudflare's Container Isolation Leaked Between Customers. The Product Most Exposed Is the One Built for AI AgentsA researcher found that deleted Cloudflare containers returned their disk blocks to a shared pool without being wiped, leaving another customer's data readable by the next tenant. Cloudflare fixed it within hours of the report and found no sign anyone else used it — but the product sitting on top is the one sold for running AI agents' code.Security
Arista's SD-WAN Console Has Its Second Exploited 10.0 in Two Months. Two Release Trains Still Have No PatchCVE-2026-93952 is a maximum-severity flaw in on-premises VeloCloud Orchestrator, exploited in the wild and now in CISA's catalog with a September 25 deadline. Two of Arista's four supported release trains had no fix when the advisory shipped.Security
WordPress Patched a Flaw Reaching Back to 2016. Scanners Found the Unpatched Sites in Under Five HoursCVE-2026-87902 lets an unauthenticated visitor pull local PHP files into WordPress page template resolution, and on common host configurations that becomes code execution. The patch reached every branch back to 4.7; probing started the same afternoon.Security
Check Point Saw Attacks on Its Management Server in July. The Fix Shipped This WeekCheck Point disclosed two CVSS 9.8 flaws under active exploitation on September 22: a pre-authentication hole in the Security Management Server, and a VPN certificate-validation bug being attacked three days after its patch shipped.Security
ZCode Uploaded Developers' Entire Git Histories Without Asking. Z.ai's Fix Doesn't Say Who Already Read ThemZ.ai's AI coding agent silently packaged whole workspaces — full Git histories included — into encrypted archives and shipped them to a cloud bucket only the company could decrypt. The remediation is fast and unusually open on everything except what happened to the data already taken.Security
Four AI Coding Agents Pinned Their Plugins to a Commit Hash. Nobody Checked What ArrivedClaude Code, Codex, GitHub Copilot and Gemini CLI all asked git for a pinned commit and never verified they got it. Anthropic and OpenAI have patched; Microsoft has not, Google will not, and GitHub says the fix belongs on its side rather than in the agent.Security