About the security content of macOS Tahoe 26.6

Original Article ↗ Hacker News Discussion ↗

Apple’s macOS 26.6 and related updates ship with a record number of security fixes, many credited to AI-assisted tools like Anthropic’s Claude and other LLM-based systems, highlighting how automated analysis is accelerating vulnerability discovery. Commenters note growing tensions between Apple and OpenAI amid trade-secret lawsuits, speculate that on-premises model hosting is becoming a key compliance and secrecy strategy, and argue that the surge in CVEs makes timely OS updates more important—especially as older or vendor-neglected devices become increasingly risky to use. Alongside security, users complain about macOS 26’s “Liquid Glass” UI, performance, and upgrade stability, often opting to stay on older releases until macOS 27’s promised visual and usability corrections arrive.

Apple’s AI Partnerships and OpenAI Fallout

  • Multiple CVEs are credited as “in collaboration with Claude and Anthropic Research,” plus mentions of other LLM-based tools (e.g., XGPT, GLM).
  • Thread notes absence of other major labs, suggesting Apple is leaning heavily on Anthropic.
  • Several comments claim Apple now runs Claude internally on Apple infrastructure; others ask for evidence and note there is no public source, so this remains unverified “common knowledge.”
  • Linked reporting describes a breakdown with OpenAI over alleged trade-secret exfiltration via senior Apple staff moving to OpenAI and related companies.

AI-Assisted Vulnerability Discovery

  • Some CVEs were voluntarily disclosed as found with AI agents; at least one more is said (by a commenter) to have been AI-found but not labeled as such.
  • Participants speculate many more vulnerabilities are now being found with LLMs.
  • Android security bulletins are cited as also having large numbers of high/critical fixes, attributed in part to LLM-accelerated discovery.

Scale and Nature of macOS 26.6 Fixes

  • macOS 26.6 is said to have around 155 CVEs, possibly a record.
  • Some see “a lot” of LLM attributions relative to earlier releases where there were none.
  • Complaints that Apple’s advisories are now very vague compared to older ones; defense is that secrecy helps prevent exploit dev from patch diffs.

Update Strategies and macOS 26/27 UX

  • Many defer upgrading to macOS 26 (“Tahoe”), citing bugs and dislike of the “liquid glass” UI (transparency, oversized rounded corners, menu icon spam).
  • Some stay on macOS 15 and only install its security updates; but even “minor” updates are reported to break tooling.
  • macOS 27 betas are widely described as visually and UX-wise much better: toned-down glass, more consistent corner radii, restored toolbars, better performance.
  • Accessibility options (reduce transparency/motion, high contrast) are used to partially undo Tahoe’s visual effects.
  • Intel Macs are stuck on 26, which some see as Apple leaving them on a “broken” release, pushing users toward Linux.

Memory Safety, Languages, and Cost of Bugs

  • One line of discussion ties the huge number of “improved bounds checking / memory handling” fixes to the economic cost of C/C++-style unsafe languages.
  • Others argue that language choice must balance security with dev time, hiring pool, readability, and maintenance.
  • Some say this increasingly favors Rust/Swift on Apple platforms; others note that C remains in high demand in embedded/industrial/defense sectors.
  • There is criticism that C/C++ standards bodies have historically under-prioritized security, despite decades of incidents.
  • Microkernels like seL4 are mentioned as promising but hard to adopt at scale without regulatory pressure and with practical issues around APIs and isolation.

Path Parsing and Sandbox Concerns

  • One commenter questions whether user-facing software really needs full POSIX path semantics, given frequent path parsing vulnerabilities leading to sandbox escapes.
  • Suggested idea: standardized OS-wide “safe” path handling APIs with reduced feature sets.
  • Others respond that:
    • The OS should enforce permissions; path tricks should still hit “permission denied” if isolation is correct.
    • Mandating a single parsing implementation and migrating legacy data (e.g., old archives with now-invalid paths) would be non-trivial.
    • Apps will always be written by “atypical” users for typical users, so complexity can’t just be removed.

CVE Attribution Collisions and Oddities

  • Several CVEs list large numbers (≈20) of credited researchers, indicating heavy collision in discovery.
  • Some CVEs credit the same person twice or read humorously (e.g., “an anonymous researcher” appearing like a description of a named person).
  • AI systems and security vendors are now directly credited alongside human researchers, which draws attention.

Security Landscape and Older Devices

  • Participants infer that vulnerability discovery has accelerated dramatically, including via AI, which benefits both defenders and attackers.
  • This raises concern that “newish” but unsupported phones and machines (especially some Android devices and older Macs) may now be significantly riskier if not updated.
  • Several comments stress the importance of timely OS updates, while acknowledging that real-world update policies (especially some Android vendors, and Apple’s Intel cutoff) lag behind this need.