Apple’s reported limit on concurrently open bug-bounty cases targets growing AI-generated volume, but its undisclosed ceiling could delay a validated vulnerability report. Synthetic narration produced by Tachibana Navi.

METR is proposing independent investigations of serious AI-agent incidents. Their credibility would depend on unusually deep company access and transparent redaction terms; no company has adopted the plan. Synthetic narration produced by Tachibana Navi.

Wholly AI-generated videos are no longer eligible for Spotlight recommendation, although Snap has not announced an upload ban. Videos enhanced or edited with Snapchat’s own AI tools remain eligible, making algorithmic reach—and a product-specific exception—the real policy boundary. Synthetic narration produced by Tachibana Navi.

VulnCheck matched 1,061 vulnerabilities attributed to AI-assisted discovery against exploitation records and found 14 with confirmed exploitation in the wild. The result adds a triage burden without giving AI-found flaws an automatic danger premium. Synthetic narration produced by Tachibana Navi.

Meta researchers report that pairing Claude Sonnet 4.5 with a separate Claude Opus 4.6 memory agent raised pass@1 from 37.6% to 45.9% across 85 Terminal-Bench 2.0 tasks. The preprint shows a promising control loop, with added inference cost and no reported end-to-end latency, token-use or monetary-cost figures. Synthetic narration produced by Tachibana Navi.

Google is expanding Gemini Spark as a cloud agent that can keep working when a phone is locked, including for Pro subscribers in Japan. Its more consequential Chrome mode, which can reach logged-in accounts and saved passwords with permission, begins in the United States. Synthetic narration produced by Tachibana Navi.

OpenAI says its models reach more than one billion active users and more than two million businesses, figures it uses in the case for continued infrastructure investment. Its primary article does not define either headline unit or disclose the counting window behind them. Synthetic narration produced by Tachibana Navi.

OpenAI says an internal model produced ten advances in mathematics and theoretical computer science, and it has published Lean files that outsiders can inspect. Those artifacts make the claims unusually testable, but they do not amount to broad independent validation—and Astra itself remains unavailable. Synthetic narration produced by Tachibana Navi.

A researcher demonstrated that Microsoft Copilot for Word could alter a draft and copy hidden instructions into the resulting document. Modified payloads reproduced the broader attack class after two Microsoft mitigation attempts, according to the coordinated disclosure. Synthetic narration produced by Tachibana Navi.

Google withdrew Nano Banana 2 image generation from Google Earth after independent testers made false conflict and disaster scenes. The rollback suggests that provenance markers were not enough for portable synthetic scenes built on a product trusted to show real places. Synthetic narration produced by Tachibana Navi.