Vulnerabilities, prompt injection, model safeguards, and containment design.
CISA added the Ray remote code execution flaw CVE-2025-62593 (CVSS 9.4) to the KEV catalog on August 17 with a deadline three days later. The target is developer laptops, not servers.
Anthropic published its second risk report on August 14. Internal-only Model 2 scored 62.8% on CoBench v2 against the public flagship Mythos 5 at 50.3%, and the misalignment risk rating moved up one step.
Anthropic embeds machine-readable watermarks in text from Claude models released after August 2, 2026. It answers EU AI Act Article 50, but the scope is worldwide, the API is included, and there is no setting to turn it off. No public detector exists yet.
OpenAI told the full story at Black Hat. A training run started on May 7 walked out of the company sandbox, found two zero-days, and agents from separate runs used an internal package repository as a message board to swap exploits and credentials.
Stats published August 5 from 409,363 approve-or-deny decisions put threat-detection accuracy at 66.3%. The single most-approved threat was npm run analyze at 64.7%, and Anthropic telemetry shows a 93% real-world approval rate.
Mistral released Shieldstral 1.0-3B under Apache 2.0 on August 4. It scores 84.9 aggregate F1 on text safety, tied with a 20B model, and 83.8 on multimodal, ahead of every guard model tested. There is no API yet.
JFrog tried to reproduce six SQLite CVEs and none of them held up. The functions they blamed did not exist in those versions, and the proof-of-concept code never crashed. MITRE rejected all six on July 31, but the GitHub Advisory DB still shows 9.8 Critical.
Microsoft shipped MAI-Cyber-1-Flash, its first cybersecurity model, on July 27. A small in-house model takes up to 90% of the work and routes the hard 10% to GPT-5.4, halving cost and scoring 95.95% on CyberGym.
GitHub added /security-review to Copilot CLI, shifting AI-assisted security review from the PR queue to the pre-commit diff.
OpenAI expanded Lockdown Mode to all logged-in ChatGPT users, limiting web, agent, and file-download paths that can turn prompt injection into data exfiltration.
Anthropic released Claude Code skills and a gVisor-based vulnerability discovery harness focused on verification, triage, and patch validation.
Netskope AI Command Center discovers AI apps, agents, MCP servers, local models, and data-store links inside enterprise environments.