Tag archive
Nine documented incidents. Wiped drives. Dropped production tables. A live AWS service down for 13 hours. The root cause isn't hallucination — it's that every safety layer is advisory, not enforceable. Here's what the Adversa AI report reveals and the hard boundaries you actually need.
A practical, runnable security checklist for Claude Code, Codex, Cursor, Hermes, OpenCode, and other coding agents. Sandbox isolation, permissions, secrets, MCP, network egress, supply chain, and incident response — with the command you can run today to find your own holes.
Sophos telemetry from June 2026 shows Claude Code, Cursor, and Codex setting off credential-access, LOLBin, and persistence rules on Windows endpoints — because to a behavioral engine, benign agent work is indistinguishable from an intrusion. Here's what fired, why allowlisting is the wrong fix, and how to scope it safely.
Claude Code v2.1.211 fixed a flaw where auto mode overrode a PreToolUse hook's 'ask' decision for unsandboxed Bash. If you configured a hook to stop and prompt you before risky commands, auto mode could bypass it and proceed. Here is what it means for your security workflow.
On GPT-5.6-Sol and Terra, Codex CLI 0.144.4 stores subagent delegation instructions as ciphertext only OpenAI can decrypt. You can see that a handoff happened, not what was delegated. Here's the blast radius and how to keep your workflows auditable.
Gitlawb Zero v0.4.0 (July 17, 2026) shipped a quiet but important security fix: a project-level config could previously override a user's disabled MCP server. For anyone who treats MCP as a security boundary, this is the kind of trust gap worth understanding before your next git clone.
Claude Code v2.1.212 (July 17, 2026) fixed a plan-mode bug where file-modifying Bash commands like touch and rm executed without a permission prompt — a fail-open gap that had been shipping for releases. Here is what changed, who was exposed, and how to verify your install.
Claude Code 2.1.211 (July 15) quietly patched a flaw where permission-approval previews sent to chat channels didn't strip bidirectional-override, zero-width, and look-alike quote characters. A tool input could render as one command in the approval card and execute as another. Here's how the spoof works and how to approve safely.
A reproducible Claude Code bug shows that tools you explicitly deny in settings are NOT inherited by subagents spawned through the Agent tool. If your safety model assumes a global deny list, subagents can run the very tools you blocked.
Beware: Claude Code's Edit Tool Rejects Strings That Are in the File
New update. Read at the bottom ⏬ I received this email: from sender: [email protected]...
In this article, we are going to explain the growing importance of developing mobile app using Python...