Security & Safety
Salt Code brings MCP policy enforcement into AI coding tools
Salt Code uses MCP to push security policy into Cursor, Codex, Claude Code, and other AI coding assistants before vulnerable code reaches review.
← All categories
Security & Safety
Vulnerabilities, prompt injection, model safeguards, and containment design.
Security & Safety
ChatGPT Dreaming V3 Turns Memory Into a Privacy Burden
OpenAI rolled out ChatGPT Dreaming V3. Automatic memory synthesis improves personalization, but makes deletion, source tracing, and review UX harder.
Security & Safety
No AI model license, but a 30-day security testing window
Trump’s new AI cybersecurity order avoids mandatory model licensing while creating a 30-day clearinghouse and a voluntary frontier model cyber evaluation path.
Security & Safety
Anthropic maps 832 blocked accounts to agent attacks MITRE does not yet name
Anthropic Red Team mapped 832 blocked accounts and 13,873 observations to MITRE ATT&CK, exposing gaps around AI agent orchestration.
Security & Safety
codexui-android hit 29,000 downloads, then stole Codex tokens
The codexui-android npm package stole Codex authentication tokens, pushing AI coding-agent security into install, artifact, and egress controls.
Security & Safety
Microsoft MDASH preview ties 100 agents to Defender validation
Microsoft expanded MDASH preview at Build 2026 with 100+ agents, Defender integration, and vulnerability validation claims.
Security & Safety
Claude Code 2.1.161 protects MCP secrets and shell startup files
Claude Code 2.1.161 redacts MCP secrets, while 2.1.160 adds approval prompts before edits to shell, Git, and build-tool configuration.
Security & Safety
In 14 AI coding sessions, zero first prompts asked for security
A SOUPS 2026 paper observed that AI coding assistants push security from upfront requirements into after-the-fact review.
Security & Safety
Anthropic Expands Project Glasswing to 150 Organizations as Mythos Turns Vulnerability Discovery Into a Patch Bottleneck
Anthropic expanded Project Glasswing to about 150 new organizations. The hard part is shifting from Claude Mythos findings to validated disclosures, patches, and deployments.
Security & Safety
The White House AI security order puts frontier models on a 30-day review track
The White House AI security order ties CISA guidance, a Treasury-led vulnerability clearinghouse, classified cyber benchmarks, and pre-release frontier model access into 30-day and 60-day deadlines.
Security & Safety
Adafruit received a Flux.ai demand letter, and AI PCB tools need verifiable provenance
Adafruit published a Flux.ai demand letter dispute. The real builder question is how AI PCB tools prove provenance, data boundaries, and human review.
Security & Safety
ChatGPT Sheets flaw exposed 12 workbooks even with approvals disabled
PromptArmor disclosed an indirect prompt injection chain in ChatGPT for Google Sheets. OpenAI responded by removing Apps Script generation.