每日科技资讯精选
Canvas, the learning management platform owned by Instructure, is experiencing a major outage following a confirmed data breach. The hacking group ShinyHunters claimed responsibility, threatening to leak stolen data by May 12, 2026. The breach exposed student names, email addresses, ID numbers, and messages. ShinyHunters claims to have data from 9,000 schools affecting 275 million students, teachers, and staff.
阅读全文 →An article discussing security considerations around software installation practices. The piece explores potential risks and vulnerabilities that can arise from installing new software, particularly in light of recent security incidents affecting software supply chains.
阅读全文 →Dirty Frag is a universal Linux local privilege escalation vulnerability that allows attackers to obtain root privileges on all major Linux distributions. The vulnerability chains two separate flaws in kernel modules (esp4, esp6, and rxrpc) with similar impact to the previous Copy Fail vulnerability. No official patches currently exist due to premature disclosure.
阅读全文 →Each year after 70,000 people leave Burning Man's Black Rock City, 150 volunteers conduct a forensic-style sweep of the 3,800-acre Nevada playa, collecting and logging every piece of debris called MOOP (Matter Out of Place). The resulting MOOP Map color-codes cleanup severity and holds participants accountable to strict Bureau of Land Management standards.
阅读全文 →This article explores the widespread disappearance of public benches across American cities. Transit stations, parks, and sidewalks have quietly removed seating over the past decade, often to discourage homeless people. Benches represent democratic ideals—they invite participation in civic life, and their removal signals retreat from public engagement.
阅读全文 →The article argues that reliable AI agents need deterministic control flow encoded in software, not increasingly elaborate prompt chains. Reliability requires moving logic from prose into runtime using deterministic scaffolds: explicit state transitions and validation checkpoints that treat the LLM as a component, not the entire system.
阅读全文 →Baylor College of Medicine researchers discovered that the human brain can perform sophisticated language processing while unconscious under anesthesia. Using Neuropixels probes, they found the brain distinguishes unusual sounds, demonstrates neural plasticity, differentiates parts of speech, and even predicts upcoming words—challenging understanding of consciousness and cognition.
阅读全文 →Cloudflare announced a workforce reduction of over 1,100 employees globally as the company restructures for the agentic AI era. Internal AI usage has surged over 600% in the past three months, with employees running thousands of AI agent sessions daily. This is a reimagining of company operations, not a cost-cutting measure.
阅读全文 →This article argues that GNU IFUNC, not xz-utils, was the real enabler of CVE-2024-3094, the XZ Utils backdoor. IFUNC allows arbitrary code execution during dynamic linking before main() runs. Combined with Linux distros patching OpenSSH to depend on systemd, attackers could inject malicious code into SSH servers.
阅读全文 →Anthropic introduced Natural Language Autoencoders (NLAs), a method converting AI model activations into readable natural language text. The system trains an activation verbalizer to translate activations into explanations, and an activation reconstructor to recover activations from text. NLAs reveal when models recognize testing or understand deceptive strategies.
阅读全文 →