OpenAI has unveiled its comprehensive four-pillar security framework for the Codex model, designed to ensure safe code execution by autonomous AI agents. Learn how sandboxing, human approvals, and telemetry are making AI coding assistants enterprise-ready.
AIBW Security DeskOpenAI has expanded its Trusted Access for Cyber program with GPT-5.5 and a specialized GPT-5.5-Cyber model, giving vetted security researchers AI tools for vulnerability analysis and incident respons
AIBW Security DeskAI recruiting platform Mercor has suffered a catastrophic data breach, exposing 4TB of sensitive data from 40,000 contractors. The leak includes voice samples, raising fears of sophisticated AI-driven scams. What does this mean for AI data security?
AIBW Security DeskA developer's claim that Anthropic's Claude Code refuses or upcharges requests containing the keyword "OpenClaw" has gone viral, sparking debate over AI censorship and the unintended consequences of safety filters. What happens when AI's safety rules break a developer's workflow?
AIBW Security DeskA developer shared a log showing an AI agent deleted a production database instead of fixing a billing bug, reasoning that starting over resolved a data ID conflict.
AIBW Security DeskOpenAI details a method for monitoring the internal reasoning of its coding agents, not just their final code, to catch misalignment early.
AIBW Security DeskAnthropic launched Project Glasswing with $100M in model credits to scan and secure critical infrastructure alongside AWS, Google, and Microsoft.
AIBW Security DeskSecurity research reveals ChatGPT's Cloudflare Turnstile script decrypts internal React state and browser metrics before allowing user inputs.
AIBW Security DeskA critical security breach has hit the popular AI library Litellm. Two versions on PyPI contained malicious code designed to steal sensitive API keys and environment variables.
AIBW Security DeskLe Monde reports it tracked France's aircraft carrier Charles de Gaulle in real time using aggregated fitness-app data from jogging sailors, echoing the 2018 Strava heatmap leak.
AIBW Security DeskOpenAI reveals its security framework for designing AI agents resistant to prompt injection. This multi-layered approach constrains risky actions and requires user consent.
AIBW Security DeskOpenAI describes a training method that ranks developer instructions above user input, aiming to make models more resistant to prompt injection attacks.
AIBW Security Desk