Daily AI brief

Copilot adds enterprise sandbox controls; new reports examine learning and cyber espionage

GitHub expands JetBrains controls and context, ByteDance introduces EdgeBench, and Anthropic describes an AI-orchestrated espionage campaign.

2026-09-09·Updated 2026-09-09·Primary sources shown for every item

01Products
GitHub adds enterprise sandbox policies to Copilot for JetBrains
GitHub updated Copilot for JetBrains with support for enterprise-managed sandbox policies and enterprise policy diagnostics. The release also adds global project context in chat and cross-file cursor jumps for next edit suggestions.
Why it mattersTeams managing Copilot inside JetBrains can assess policy controls alongside coding assistance in the same integration. Project context and cross-file navigation also affect how developers work across a codebase.
Read analysis
Related directory entriesGitHub Copilot — The most widely deployed coding assistant, available in most major editors.
SourcesGitHub Changelog (opens in a new tab)
02Research
ByteDance introduces EdgeBench to measure real-world environment learning
ByteDance Seed introduced EdgeBench as a way to measure real-world environment learning. The team also claims a new scaling law; its announcement gives neither the relationship's form nor numerical results.
Why it mattersFor builders choosing systems that must learn from their operating environment, the evaluation target is relevant. Whether EdgeBench can inform that choice depends on what it measures and how closely its tasks match deployment conditions.
Read analysis
SourcesByteDance Seed (opens in a new tab)
03Policy & safety
Anthropic reports disruption of an AI-orchestrated cyber espionage campaign
Anthropic published an account of disrupting what it describes as an AI-orchestrated cyber espionage campaign. It characterizes the attack as highly sophisticated, but the announcement does not specify the systems involved or the intervention used.
Why it mattersSecurity teams need to distinguish AI assistance from AI coordination when deciding which activity to detect and investigate. A documented incident could help make that distinction, provided the account explains the division of work between people and AI.
Read analysis
SourcesAnthropic (opens in a new tab)

GitHub adds enterprise sandbox policies to Copilot for JetBrains

ByteDance introduces EdgeBench to measure real-world environment learning

Anthropic reports disruption of an AI-orchestrated cyber espionage campaign