Editorial: Two themes thread today's stories: convenience that increases risk, and cultural shifts that rewrite incentives. A tiny kernel bug and big LLMs both illustrate how small technical or social changes cascade into outsized consequences.
In Brief
I believe there are entire companies right now under AI psychosis
Why this matters now: Companies that treat large language models as authoritative risk making product, security, and legal decisions based on hallucinations—exposing customers and revenue to harm this quarter.
Mitchell Hashimoto's blunt post on X kicked off a broad thread about what happens when teams stop questioning outputs from LLMs. Read the original post and the discussion for lots of concrete failure modes: support tickets containing nonexistent JSON fields, product managers accepting fabricated API contracts, and leadership that prefers model output over engineering checks.
"I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them."
Hacker News commenters converged on practical guardrails: treat model-generated text as hypotheses, not facts; keep humans in the loop for validation; encode decisions into specs and tests. The cultural lesson is immediate: convenience without verification is an operational hazard, especially where customers or compliance are involved.
Project Gutenberg — keeps getting better
Why this matters now: Project Gutenberg's UX and tooling improvements make a growing archive of public-domain books easier to reuse—important for researchers, TTS projects, and legal-free digital libraries right now.
The volunteer-run Project Gutenberg quietly refreshed its site and continues to expand a collection now billed as "over 75,000 free eBooks"—including EPUBs, Kindle files, LibriVox human audiobooks, and machine-generated audio from a recent collaboration. The maintainers joined the conversation on Hacker News noting steady work on UX and bug fixes; readers asked for better provenance, OCR workflows, and metadata that makes the site more useful for downstream projects. See the foundation's homepage for details at Project Gutenberg.
The sigmoids won't save you
Why this matters now: Arguing that AI progress must plateau because “exponentials become sigmoids” is a weak defense unless you can point to a clear mechanism for the slowdown.
The essay reframes a common anti-alarmist talking point into a call for explicit models and honest uncertainty accounting. The author suggests Lindy’s Law as a better default—if something has been accelerating recently, it plausibly keeps accelerating—and warns against laundering ignorance into neat curves. The piece pushes technologists and policymakers to demand evidence for any claim that the current growth spurt will stop imminently; read the full argument at Astral Codex Ten.
Deep Dive
A 0-click exploit chain for the Pixel 10
Why this matters now: A trivial kernel-escalation bug in the Pixel 10 VPU driver allowed attackers to map arbitrary physical memory from userspace—meaning remote compromise could yield full device root with minimal code.
Google's Project Zero rebuilt a zero-click chain initially used on Pixel 9 and adapted it to Pixel 10, only to find a second, shockingly simple privilege-escalation: the VPU device exposed MMIO and let userspace mmap far beyond the intended register region. According to the report, "by specifying a size larger than the register region in an mmap syscall, the caller can map as much physical memory as they want into userland," and because the kernel lives at a fixed physical offset on these devices, an attacker can overwrite kernel code.
"Achieving arbitrary read-write on the kernel with this vulnerability required 5 lines of code and writing a full exploit for this issue required less than a day of effort."
There are three immediate takeaways. First, auto-decoding features and convenience parsers increase attack surface—thumbnailing, previewing, or auto-processing incoming media can put complex third-party decoders into privileged paths. Second, driver hygiene matters: small mistakes in MMIO handling or bounds checking create catastrophic capability leaks. Third, Google’s triage timeline was notably quick here—the bug was reported in November and patched in February—showing improvement in response, though the broader Android ecosystem still suffers fragmentation in updates.
For anyone building embedded or driver-level code: avoid exposing raw physical or MMIO regions to untrusted contexts, enforce strict bounds, and assume that any convenience decoding can be an exploitation vector. The Pixel 10 case is a neat, alarming illustration that one misplaced mmap can undo modern kernel defenses.
Frontier AI has broken the open CTF format
Why this matters now: Frontier LLMs are reliably solving many CTF challenges, turning competitions into money-and-agent orchestration contests rather than human skill ladders.
The author argues that models like Claude Opus 4.5 and GPT-5.5 can "one-shot" many previously meaningful CTF problems, collapsing a learning pipeline that introduced novices to progressively harder tasks. The post argues that leaderboards now reward who can spin up the best agent or burn the biggest token budget, not who wrote the clever exploit.
"CTFs feel much more like a cheesable mess than a competition"
This is an incentive collapse with tangible downstream effects. Beginners lose the gradual feedback loop that teaches debugging and exploit thinking. Challenge authors stop investing time in craft if their puzzles are trivially solved by a model. Hiring signals degrade: a top score may reflect orchestration or compute budget, not human expertise.
The community responses are instructive: suggested fixes include dual leaderboards (human-only vs. AI-augmented), moving high-stakes rounds offline or in-person, or redesigning problems to test long-running skills or creative proof-of-work that’s hard for stateless models to fake. Practically, contest organizers now have a choice: accept the new meta (and make CTFs about orchestration), or redesign to preserve human learning pathways—either direction reshapes how the security talent pipeline develops.
Closing Thought
Convenience—whether a tiny mmap, a model that writes convincing JSON, or an LLM that solves puzzles—accelerates value and risk together. The hard work now is less about inventing tools and more about designing incentives, boundaries, and verification that keep systems safe and cultures honest.