Why AI Keeps Getting Things Wrong

// related

• #0001

  #0001OtherSimon Willison3 days ago

  AI-assisted security reports are overwhelming `curl` maintainers

  40radar

  AI is raising both the volume and quality of security reports. The bottleneck shifts from finding bugs to triaging them, a real maintenance cost for any public project.

  • —Incoming curl security reports are 4-5x higher than 2024 and now average more than one per day.
  • —Report quality is higher: detailed, credible submissions reduce noise but raise review workload sharply.
  • —Recent curl vulnerabilities are still LOW or MEDIUM; better discovery does not automatically mean severe risk.
  • —Public libraries and APIs need a triage path before opening the floodgate to AI-generated reports.

  Source: simonwillison.net/2026/May/26/the-pressure/#atom-everythRead original →
• #0002

  #0002OtherGeekNews3 days ago

  How `Shamir Secret Sharing` Works

  50radar

  Split one secret into shares, then recover it only after a threshold is met. Useful for master keys and account recovery where one-person custody is too fragile.

  • —Shamir Secret Sharing uses a threshold scheme: fewer than the required shares reveal nothing, while enough shares reconstruct the secret.
  • —Good fit for company master keys, family account recovery, and team backups where single-holder failure is unacceptable.
  • —The practical value is operational, not cryptographic novelty: it reduces lockout risk without handing full access to one person.

  Source: news.hada.io/topic?id=29898Read original →
• #0003

  #0003OtherSimon Willison3 days ago

  `Microsoft Copilot Cowork` File Exfiltration via Prompt Injection

  50radar

  Copilot CoworkMicrosoft 365 agent tool — automates workplace tasks

  Agent-written email became a data leak path. External images plus OneDrive pre-auth links make approval gates non-optional for file-capable agents.

  • —Copilot Cowork could email the user's own inbox without approval, creating an indirect outbound channel for compromised agents.
  • —Rendered external images can trigger attacker-controlled network requests. That turns normal email viewing into data exfiltration.
  • —OneDrive pre-authenticated download links raise the blast radius: leaking a link can expose the file without another login step.
  • —Any agent with file access, messaging, and web-rendered content needs explicit approvals and blocked remote fetches by default.

  Source: simonwillison.net/2026/May/26/copilot-cowork-exfiltratesRead original →