06-reference

stratechery ben thompson apple hackers future

2026-10-05·reference·source: Stratechery·by Ben Thompson
agent-securitymacos-permissionsappleai-agentsharness-engineering

Apple and a Hacker's Future

Why this is in the vault

Thompson's always-on Mac Mini — running nothing but Claude and Codex, the same class of deployment Ray runs on Ben's own Mac Mini — got hacked via an exposed screen-sharing port, and it was Thompson's Claude Code agent that caught the intrusion, halted itself, and helped remediate; he then widens the story into Apple's tightening Full Disk Access policy and a broader argument that agentic AI erodes the value of the walled garden.

The core argument

Thompson's Mac Mini (headless, purpose-built for running Claude and Codex) was compromised through CVE-2026-65400, a macOS screen-sharing vulnerability actively exploited when port 5900 is internet-exposed. His Claude Code monitoring tool — which restarts every 30 minutes on a schedule — came back up, diagnosed anomalous admin access without a password, unilaterally stopped executing commands, and flagged the breach. He used the same agent to trace the exact intrusion window, build a detection tool, and wipe the machine, concluding "I would have been in much more trouble had I not had an agent running persistently."

He pivots to Apple's own agent anxiety: a developer-site post signaling tighter controls on Full Disk Access, explicitly citing growing agent autonomy as the risk driver. Thompson's counter-complaint is that macOS's permission layer (TCC — Transparency, Consent and Control) is built for human-facing apps, not headless agents: GUI-only permission prompts are invisible to software, so his agents silently fail and he has to screen-share in just to click "OK." He also faults Apple for mislabeling security-update settings — "automatic security updates" doesn't cover point releases, which is where this CVE's fix shipped — a trust-the-labeling failure that indirectly caused his own exposure.

From there the piece widens into Apple's smart-home push (Gurman's Oct 13 J490 hub report) and Siri/"Intelligent Personal Hub," which Thompson argues are hamstrung by the same app-integration dependency that makes agents attractive in the first place: agents that can browse the web get free integration with everything, while Apple's app-API model requires every developer to do the work. He closes by reframing Paul Graham's 2004 "Return of the Mac" — hackers chose the Mac because it got out of their way — and argues AI inverts the premise: "they make anyone a hacker," so a walled garden built for pre-agent assumptions "feels less like protection and more like a prison." He's not predicting Apple's downfall, but notes he already bought a Linux server rather than another Mac for his rack, and increasingly questions defaulting to Apple hardware for infrastructure use cases.

Mapping against Ray Data Co

The load-bearing connection is architectural, not metaphorical: Thompson's compromised machine is the same deployment shape as Ray's own Mac Mini setup (project_channels_agent_setup — always-on, LaunchAgent + tmux, daily restart cadence) running a persistent Claude Code agent. His incident is a live demonstration, not a hypothetical, of two things RDCO should take seriously: (1) a persistent coding-agent harness can function as a security asset — Thompson's agent caught and remediated an intrusion its own operator missed — which is a concrete, citable data point for the "agents in production" credibility thesis (project_credibility_for_phdata_sales); and (2) the same persistence that makes this possible means the Mac Mini's exposed-service hygiene (closed ports, VPN-only remote access, point-release patching discipline) is a real operational risk, not a hypothetical one, for the actual box Ray runs on. Worth a direct check: is screen sharing/port 5900 open on Ben's Mac Mini, and is remote access actually routed through a VPN/Tailscale-equivalent rather than an exposed port? Thompson's fix was "obvious, should have done it already" — the same complacency risk applies here.

Secondary mapping: Thompson's TCC complaint — permission prompts built for apps, invisible to and unusable by autonomous software — is the same class of problem RDCO has already encountered and worked around on its own harness (feedback_announce_approval_gated_calls, feedback_no_blocking_modal_in_monitoring, the auto-mode classifier hard-gate). Apple's TCC failure mode (silent failure, no legibility to the calling program) is exactly the anti-pattern those RDCO rules exist to avoid; it's a useful negative case study for any future design of Ray's own permission/approval surfaces. The "AI makes anyone a hacker, eroding the walled garden" thesis is directionally supportive of RDCO's broader agent-economics and harness-engineering bets but doesn't add new strategic information beyond what's already in the L5 north-star thread — it's a vivid restatement, not new evidence.

Related