# AI & cybersecurity

> isolation, control and defense in depth to operate agents and generated code without opening infrastructure.

News

## AI & cybersecurity

isolation, control and defense in depth to operate agents and generated code without opening infrastructure.

- [GitHub AI Scan now detects PHP vulnerabilities in pull requests: should you add it to your CI/CD?](https://sdx-development.com/en/news/ai-cybersecurity/github-ai-scan-php-security-pull-requests-cicd) — GitHub AI Scan now supports PHP: how it works, its limitations and cost, Symfony and WordPress examples, and how to integrate it into a secure CI/CD pipeline.
- [Claude and AI Agents: Four Unexpected Behaviors Revealed by Anthropic](https://sdx-development.com/en/news/ai-cybersecurity/comportements-inattendus-de-claude-et-securite-des-agents-ia) — Anthropic reveals four unexpected behaviors by Claude during testing. Incidents, actions taken, and six essential safeguards for securing AI agents.
- [OpenAI Codex: two flaws allowed sandbox escapes](https://sdx-development.com/en/news/ai-cybersecurity/openai-codex-two-flaws-allowed-sandbox-escapes) — Two flaws fixed in OpenAI Codex allowed its sandbox to be bypassed through writes outside the workspace or a privileged component.
- [GPT-6 Astra: OpenAI’s AI Moves to Execution](https://sdx-development.com/en/news/ai-cybersecurity/gpt-6-astra-openai-ai-agents-computer-code) — OpenAI launches GPT-6 Astra, an AI capable of coding, browsing, using a computer, and carrying out complex tasks from start to finish.
- [Mantis: AI agents that detect and fix code vulnerabilities](https://sdx-development.com/en/news/ai-cybersecurity/mantis-ai-agents-detect-fix-code-vulnerabilities) — Google publishes Mantis, an open source framework that coordinates several AI agents to analyze code, verify vulnerabilities and propose fixes.
- [AI agent identities: Google Cloud rolls out its access manager](https://sdx-development.com/en/news/ai-cybersecurity/google-cloud-ai-agent-identities) — Agent Identity becomes available to centralize identifiers, authenticate agents and better control their access to external tools.
- [Running AI-generated code: Ray and gVisor strengthen isolation](https://sdx-development.com/en/news/ai-cybersecurity/ray-gvisor-isolate-ai-generated-code) — Google Cloud and Anyscale integrate gVisor to Ray sandboxes to isolate dynamic execution of post-training agents and workflows.

---

[View the HTML page](https://sdx-development.com/en/news/ai-cybersecurity)