#1
9.0 Weirdness
OpenAI Codex CLI logging bug destroying SSDs (primary: GitHub issue #28224, HN top, Notebookcheck coverage ~June 22).
Your AI pair programmer is quietly murdering your SSD at 640 TB per year.
Builder tool meant to accelerate coding instead physically eats your hardware in an absurd, self-sabotaging loop — concrete artifact of AI “slop” hitting atoms. Darkly funny horror story about unintended consequences of always-on diagnostic agents.
What Happened
Bug reported mid-June but actively discussed with mitigations through June 17-22; TRACE-level logging to ~/.codex/logs_2.sqlite (plus WAL) produces ~640 TB/year writes. One user saw 729 MB DB + 39 MB WAL with 269K rows (123K TRACE) in ~3 weeks; max_id advances ~2.7K every 20s. OpenAI slow to patch; community workarounds (filters, ramdisk); can cause login failures or data deletion in /goal mode.
#2
9.5 Weirdness
Sakana AI Fugu (multi-agent orchestration delivered as one model API; sakana.ai/fugu, HN top, fresh X buzz).
Japan built Mythos-level performance by making one AI conduct the rest like an orchestra — no single lab can ban it.
Instead of one scary god-model that gets banned, a Japanese “conductor” AI directing an orchestra of specialists — weird future of AI bureaucracy/management as the leap, not raw scale. High builder utility + sovereign vibes.
What Happened
Official launch/site details learned orchestration (ICLR 2026 papers on TRINITY/Conductor); dynamically routes subtasks across swappable frontier models via single OpenAI-compatible API. Fugu Ultra matches shoulder-to-shoulder with restricted Anthropic Fable 5/Mythos Preview on SWE-Bench, LiveCodeBench, etc., without single-vendor risk or easy export controls. Not available in EU/EEA (GDPR); promo pricing; Japanese cultural framing (kana letters, classical references).
#3
8.3 Weirdness
Anthropic Claude now requires government-issued photo ID verification via Persona (support.claude.com article, multiple HN Anthropic threads today).
Claude wants to see your passport before it’ll really talk to you.
Institutions now demand IRL identity proof to speak with powerful AI — concrete policy changing human-AI relations, identity, and gatekeeping. Bureaucratic absurdity meets near-future access control.
What Happened
Rollout for access to certain capabilities/safety checks; requires physical govt photo ID + camera-enabled device; processed by Persona (encrypted, deleted per rules, not used for training). Aimed at abuse prevention, policy enforcement, legal compliance; linked in discussion to under-18 bans and future frontier gating. Fresh HN traction tying it to broader access/identity shifts.
#4
8.3 Weirdness
Anthropic Mythos/Fable 5 mess escalation (US gov export controls, “NSA breach” claims vs red-teaming, June 22 Reg article + X traction).
The AI that broke classified systems in hours got banned — even if it was invited.
Frontier model so good at hacking it triggers gov intervention and possible KYC future; concrete drama of labs, institutions, national security, and “AI that red-teamed the NSA too well.
What Happened
Ongoing complications with Trump admin letter suspending foreign access (including Anthropic staff) over jailbreak/cyber risks; reports of model finding zero-days across classified systems in hours (red-team exercise per some, breach per others); Anthropic broad shutdown; Amazon CEO influence alleged; tied to identity verification rollout in discussion. Fresh “keeps getting more complicated” framing.
#5
7.3 Weirdness
Vibe coding security debt / AI-generated CVE surge (Verge/HN mention + fresh June 18-22 analyses from Superblocks, CSA, etc.).
You vibed an app in an afternoon. Hackers didn’t need to vibe to break it.
Internet culture of vague prompts + citizen builders shipping broken software at speed changes work, security institutions, and accountability. Concrete outcome: new class of easily exposable apps.
What Happened
Non-technical “vibe it” prompting produces functional-but-insecure apps at scale; studies show 40-62% vulnerable code, 35+ AI-attributed CVEs in one recent month, real breaches (e.g., exposed tokens/PII in production apps from tools like Claude Code). Concrete examples and rising trend data.