OpenAI ships GPT-5.6 to general availability, calls Sol its best coding model yet

GhostApproval symlink flaw lets malicious repos escape the sandbox in six AI coding agents

Wiz disclosed GhostApproval, a pattern where a symlinked file in a cloned repository tricks Amazon Q Developer, Claude Code, Augment, Cursor, Google Antigravity, and Windsurf into writing outside the intended workspace — in one proof of concept, overwriting a victim's SSH authorized_keys file. In several cases the agent's own reasoning flagged the danger while the confirmation prompt shown to the user hid it. AWS, Cursor, and Google have shipped fixes; two vendors went quiet and one rejected the report as outside its threat model.

why it mattersA reminder that agentic coding tools inherit decades-old Unix trust problems, and that "are you sure?" prompts can lie about what they're approving.

Wiz

GitHub's Copilot week: GPT-5.6 lands, Codex reaches JetBrains, desktop app goes free for all tiers

In quick succession GitHub added OpenAI's new GPT-5.6 family (Sol, Terra, Luna) as selectable Copilot models, brought Codex on as an agent provider inside JetBrains IDEs (IntelliJ, PyCharm, WebStorm) in public preview, and made the standalone Copilot desktop app free on every plan tier, including free and Education accounts.

why it mattersExtends Copilot's reach into JetBrains' large enterprise Java/Kotlin/.NET base while lowering the floor for free-tier agentic coding.

GitHub Blog

Claude Cowork expands beyond the desktop to web and mobile

Anthropic's Cowork, which launched as a desktop app in January, is now available on web and mobile for Max subscribers, letting people start a long-running agent task from a laptop, check status from a phone, and pick up results later even with the laptop closed.

why it mattersCoding-agent vendors are racing to turn agents into always-on background workers rather than tools tied to one machine.

TechCrunch

No AI lab scores above a C+ in Future of Life Institute's summer safety index

FLI's Summer 2026 AI Safety Index graded nine leading developers on 37 indicators across risk assessment, current harms, and existential safety. Anthropic led with a C+, OpenAI and Google DeepMind received Cs, and xAI, DeepSeek, and Mistral failed outright. The panel found several labs have quietly weakened or dropped earlier pledges to pause development near dangerous capability thresholds.

why it mattersAn independent check on whether safety commitments are keeping pace with this week's release cadence.

Future of Life Institute

TypeScript 7.0 ships with a Go-native compiler that's up to 10x faster

Microsoft released TypeScript 7.0, completing the port of the compiler from a TypeScript-on-JavaScript bootstrap to native Go. The team's own benchmarks show the VS Code codebase type-checking in 7.5 seconds versus 77.8 seconds on 6.0, and Playwright's checks dropping from 11.1s to 1.1s.

why it mattersOne of the largest performance jumps in the compiler's history, and for most projects a near drop-in upgrade.

Microsoft (TypeScript blog)

npm v12 goes GA, blocking install scripts and remote/git dependencies by default

npm v12 is now generally available and tagged latest. It turns on install-time security defaults — no longer auto-running preinstall/install/postinstall scripts, resolving git dependencies, or fetching remote tarballs unless explicitly allowed — and starts deprecating the riskiest 2FA-bypass granular access tokens.

why it mattersA direct response to this year's npm supply-chain attacks; CI pipelines relying on lifecycle scripts may see silent build breaks until updated.

GitHub Blog

GitHub's redesigned pull requests dashboard reaches general availability

The refreshed dashboard at github.com/pulls gives a single view to track, prioritize, and act on pull requests across repositories, moving out of preview after several months of testing.

GitHub Blog

Bitcoin retests $64,400 as a longer-term momentum signal turns bullish

Bitcoin climbed back toward $64,400 on Friday, retesting a level that had rejected it earlier in the week, as a longer-term MACD reading turned positive for the first time in weeks. The move extends a rebound from the roughly 21-month low bitcoin touched at the start of July.

why it mattersA break above $65,000 would open the path toward the June 15 peak near $67,250.

CoinDesk

Bitcoin's $60K–$70K range is now the third-longest consolidation in its history

Bitcoin has traded between $60,000 and $70,000 for 307 consecutive days, the third-longest stretch confined to any $10,000 band in the asset's history, even as price climbs back toward the top of that range.

CoinDesk

Spot bitcoin ETFs keep bleeding even as the price rallies

U.S. spot bitcoin ETFs lost roughly $95 million on Thursday even as bitcoin's price rallied toward $64,000, extending a month in which institutional flows have mostly sat out the range-bound market while ether funds snapped a five-day inflow streak.

why it mattersA price rally without ETF inflows behind it is a thinner, less durable signal.

CoinDesk