Someone threw a Molotov cocktail at Sam Altman’s Russian Hill mansion at 3:45 AM. It bounced off the metal gate. Security guards put out the fire. Seventy-five minutes later, police found the same guy at OpenAI’s Mission Bay headquarters, 3 miles away, threatening to burn the building down. He’s 20 years old.
The attack came three days after Ronan Farrow published an 18-month investigation with 100+ interviews alleging Altman has a “consistent pattern” of lying. Physical violence just showed up at the door of AI’s most visible leader.
In today’s indie hacker news:
- Someone firebombed Sam Altman’s home, then walked to OpenAI HQ
- The Linux kernel published official rules for AI contributions
- CPU-Z downloads were poisoned for 19 hours while the code was fine
- GLM 5.1 landed 12 Elo from Opus at 1/10th the API cost
- Apple’s cloud chief backs open-source for 90% of AI use cases
TOP STORIES
3:45 AM MOLOTOV
Someone firebombed Sam Altman’s house and walked to OpenAI to finish the job
Daniel Alejandro Moreno-Gama, 20, threw a Molotov cocktail at Altman’s Russian Hill residence at 3:45 AM on April 10. It hit the metal gate and bounced off. No injuries. About 75 minutes later, SFPD found Moreno-Gama at OpenAI’s Mission Bay HQ, 3 miles away, threatening to “burn down the building.” He was booked on 5 charges including attempted murder.
Altman posted a blog response sharing a family photo: “I am sharing a photo in the hopes that it might dissuade the next person from throwing a Molotov cocktail at our house.” He called the New Yorker piece an “incendiary article,” then walked it back on X: “That was a bad word choice.”
The details:
- 5 criminal charges including attempted murder and arson
- The New Yorker investigation by Ronan Farrow ran 18 months with 100+ interviews, published 3 days before the attack
- OpenAI spokesperson confirmed threats were also made at SF headquarters
- Anti-AI protests escalating: OpenAI office lockdown in Feb 2026, London’s largest anti-AI march in March
Why builders care: The Luigi Mangione cultural moment is being applied to AI. If you’re building in public and vocal about AI, the cultural temperature matters.
Work from any WiFi like it's your home network. NordVPN's Meshnet runs a free private mesh between your laptop, dev box, and home server. SSH from a café without exposing a port, the way you'd use Tailscale. The paid VPN on top lets you test geo-fenced Stripe checkouts or feature flags from any country.
We get a cut if you sign up. Only added for tools we use ourselves.
AI CAN WRITE, AI CAN’T SIGN
The Linux kernel published official rules for AI contributions: code yes, sign-off never
Sasha Levin (kernel stable/LTS maintainer, NVIDIA) authored coding-assistants.rst, implementing consensus from the 2025 Maintainers Summit in Tokyo. Core rule: AI agents must not add Signed-off-by tags. Only humans can certify the Developer Certificate of Origin. New attribution format: Assisted-by: AGENT_NAME:MODEL_VERSION.
Torvalds on AI code: “There is zero point in talking about AI slop.” On the policy: “The documentation is for good actors,” since bad actors won’t disclose anyway.
The details:
- 40M+ lines of kernel code, 2,134 developers contributed to kernel 6.18 (highest ever)
- One fully LLM-generated patch already merged into kernel 6.15 after human review
- Levin said he “would not make an effort to enforce the rule.” Norms will evolve organically.
- 20% of security reports to curl in 2025 were AI hallucinations
Why builders care: If you maintain an open-source project, this is your template. The DCO restriction makes it explicit: if you sign off on AI-generated code, YOU are legally certifying it.
THE LINK WAS THE MALWARE
CPU-Z downloads were poisoned for 19 hours while the actual signed code was never touched
Attackers compromised a side API on CPUID’s website to swap download links to malicious executables. The signed binaries were never touched. Code signatures checked out. The malicious downloads bundled legitimate CPUID software with a trojanized CRYPTBASE.dll, deploying STX RAT to steal Chrome stored passwords.
Exposure window: ~19 hours per Kaspersky (CPUID claims 6). Package managers like Scoop auto-downloaded the compromised installers. Same threat actor ran a fake FileZilla campaign in March.
The details:
- 5 products affected including CPU-Z v2.19 and HWMonitor v1.63 (~100M users rely on CPU-Z)
- 150+ victims identified in Kaspersky telemetry across retail, manufacturing, consulting
- C2 domain registered October 29, 2025. Months of planning.
- 20 antivirus engines flagged the ZIP on VirusTotal; 32 flagged the installer
Why builders care: Code signatures weren’t enough. If you’re distributing software, your download infrastructure is now a target, not just your build pipeline.
THE 12 ELO GAP
GLM 5.1 landed 12 Elo from Opus on Code Arena at 1/10th the API cost
Z.ai’s GLM-5.1 (754B params, MoE with 40B active, MIT license) ranks #3 on Code Arena with Elo 1529.90. Only Opus 4.6 and Opus 4.6 Thinking sit above it. First open-weight model ever in the top 3. On SWE-Bench Pro it scored 58.4, #1 overall, beating Opus (57.3).
The price gap matters most. $1.40/M input vs Opus at $15/M. 10.7x cheaper on input, 17x on output. Can run autonomously for up to 8 hours. Trained entirely on Huawei Ascend 910B chips, zero Nvidia silicon.
The details:
- Code Arena Elo 1529.90, +90 jump over GLM-5.0. SWE-Bench Pro 58.4, #1 overall.
- 200K context window (vs Opus’s 1M). Self-hosting requires ~1.49 TB and 8x enterprise GPUs.
- Available on OpenRouter ($0.95/M input), Modal (free through April), and Z.ai directly
Why builders care: A solo dev running 10M coding tokens/month pays $14 vs $150 with Opus. MIT license, no vendor lock-in. Opus still leads on hardest reasoning with 5x more context, but the price wall for serious AI coding just got demolished.
TRENDING TODAY
🌎 Apple’s cloud chief says open-source models will address 90% of use cases - Sumit Gupta (Head of Cloud and AI Infra at Apple) posted on LinkedIn that open-source models catch up “within 3-6 months.” Personal opinion, not Apple policy. 201 Reddit upvotes from a 56-reaction LinkedIn post.
📊 Qwen 3.6 community voting finalizes with 2,551 votes - The 27B dense model won at 39%, beating MoE variants. Top comment (294 upvotes): “MoE enjoyers split the vote, densocrats reap the benefits.” 580 upvotes, 254 comments.
⚠️ JSON Formatter Chrome extension (2M+ users) goes closed-source, injects adware - Went closed-source in January, then pushed “Give Freely” affiliate elements into checkout pages. Hundreds of 1-star reviews. Same week as the CPU-Z hijack.
STACK OF THE DAY
🛠 Vibescore - Grades your vibe-coded project A+ to F on code quality, test coverage, deps, and security. One command, one grade. Free, open-source, zero dependencies.
Not sponsored. We just feature tools builders would actually use.
BOOKMARKED TODAY
📈 I monitored 500 YC companies and 80% are using the same growth channel - UGC/creator content, not SEO or paid ads. Informal methodology but directionally interesting. 127 upvotes, 40 comments.
💰 $50k selling lifetime deals, my worst mistake - 340 lifetime licenses at $149 ($50,640). Same users at monthly pricing would have been ~$93K. Lifetime customers submit 3x more support tickets. 85 comments on 89 upvotes, strong debate.
📖 Steve Blank: Nowhere is safe - The lean startup godfather on how cheap drones changed military defense calculus. Not startup-relevant, but worth the read. 150 HN points.
Curated by AI, built by a human. Get this daily: indiehacker.news | X | Telegram