小红书大模型推理团队联合北京大学、上海交通大学提出 HYPIC，这是首个在混合注意力大模型上实现位置无关缓存的服务系统。在 4 个生产级混合注意力模型、5 个工作负载上，HYPIC 将首 token 延迟（TTFT）平均降低 3.25 倍、同 SLO 下可持续 QPS 提升 1.66 倍，而任务 ...

ACR Stealer campaigns use ClickFix lures, JPEG steganography, and WebDAV to steal browser tokens, passwords, PDFs, and synced ...

业界为此发展出两条降本路线。一条是位置无关缓存（PIC），它放开严格的前缀约束，让每个语义独立的片段只缓存一次、并能拼接到任意前缀之后，刚好贴合RAG与Agent的prompt组装方式，其本质可归结为沿token轴直接拼接的splice、以及重算少量token恢复跨段上下文的correction两个原语。另一条是混合注意力模型，用线性注意力替换大部分全注意力 ...

Chinese AI company Moonshot AI has launched Kimi K3, its most capable model so far and the world’s first open 3T-class AI ...

曾要4年的项目，竟被Claude Code硬生生压到两周！ 成本从300万美元，砍到十几万。 最近，Anthropic首次公开了CC大规模代码迁移「全流程」，六步框架让全网震撼。 故事的主角，是Bun之父Jarred Sumner。

From late April 2026 to mid-June 2026, Microsoft Defender Experts observed increased ACR Stealer activity across customer ...

CVE-2026-59208, a JWT matching flaw that could log users into another issuer’s account on affected multi-issuer Enterprise ...

Anaconda has acquired Kilo Code, an open-source, model-agnostic platform that embeds AI agents into the tools developers use to build software. The deal brings Kilo’s community of more than three ...

Kilo grew from zero to over 3 million developers in sixteen months because developers recommended it to other developers, drawn to its open source roots and the seamless access to the latest AI models ...

GitHub Copilot VS Code browser tools are now generally available and enabled by default for all paid Copilot subscribers. The ...

According to Socket, malicious payment SDK packages on npm and PyPI are harvesting developer credentials and CI/CD ...

Cybersecurity firm Novee has identified a GitHub Actions workflow-chain risk that can expose secrets even when checks pass, ...