We shouldn’t fear AI-generated code. We should fear “low-cost” systems using AI as an excuse to bypass decades of safety engineering. Code generation isn’t new. Aerospace and automotive have used model-based code generation for 20+ years—even on systems that reached Mars. The rule never changed: Generated code is still verified code. Use AI to write code 100x faster. Great. But review it. Trace it. Test it. Verify it. Validate it. No silent AI code. No untraceable behavior. No hidden backdoors in a safety-critical baseline. AI can make software cheaper to write. It does not make safety cheaper to prove.

What’s one technology everyone is excited about that you think is completely overhyped?

What’s something everyone seems to want that you genuinely don’t care about?

Funny how AI “hacking” gets framed as research when the target is a government system. Try the same thing against a private company. Now it’s unauthorized access, preserved logs, subpoenas, discovery, damages, and lawyers asking who approved it. Governments debate. Private companies litigate. 😂 That may be the fastest AI safety mechanism we have.

Don’t complain about AI accessing your data. Block it. OpenAI, Anthropic, and other AI services still rely on networks, domains, IPs, and infrastructure you control. If your company is serious about protecting sensitive data, enforce it at the firewall. We’re happy to assist companies with securing their data, controlling AI access, and building stronger network boundaries.

Want to “turn off AI” on your network? You can. AI is not magic. It still needs DNS, IP addresses, TCP/IP, data centers, power, chips, and network access. Here is what blocking OpenAI and Anthropic actually looks like at the firewall. ANTHROPIC / CLAUDE Anthropic publishes fixed network ranges. Inbound IPv4: 160.79.104.0/23 Inbound IPv6: 2607:6bc0::/48 Anthropic outbound traffic: 160.79.104.0/21 That /21 covers: 160.79.104.0 through 160.79.111.255 If your firewall supports CIDR rules: DENY 160.79.104.0/21 DENY 2607:6bc0::/48 For outbound client access, also block the service domains: api.anthropic.com *.anthropic.com claude.ai *.claude.ai Now OpenAI. This gets more interesting. OpenAI publishes multiple IP lists depending on what is connecting to you: GPTBot 132.196.86.0/24 172.182.202.0/25 172.182.207.0/25 20.125.66.80/28 20.171.206.0/24 20.171.207.0/24 4.227.36.0/25 52.230.152.0/24 74.7.175.128/25 74.7.227.0/25 74.7.227.128/25 74.7.228.0/25 74.7.230.0/25 74.7.241.0/25 74.7.241.128/25 74.7.242.0/25 74.7.243.128/25 74.7.244.0/25 OAI-SearchBot currently includes: 104.210.140.128/28 13.66.216.176/28 135.234.64.0/24 172.182.193.224/28 172.182.193.80/28 172.182.194.144/28 172.182.194.32/28 172.182.195.48/28 172.182.209.208/28 172.182.211.192/28 172.182.213.192/28 172.182.224.0/28 172.203.190.112/28 172.203.190.128/28 172.203.190.80/28 20.14.99.96/28 20.168.18.32/28 20.169.6.224/28 20.169.7.48/28 20.169.77.0/25 20.171.123.64/28 20.171.53.224/28 20.25.151.224/28 20.42.10.176/28 23.102.145.48/28 4.227.36.0/25 40.67.175.0/25 40.90.214.16/28 51.8.102.0/24 74.7.175.128/25 74.7.228.0/25 74.7.228.128/25 74.7.229.0/25 74.7.229.128/25 74.7.230.0/25 74.7.241.128/25 74.7.242.128/25 74.7.243.0/25 74.7.244.0/25 OAI-AdsBot: 130.131.229.128/25 20.168.90.128/25 OpenAI also maintains separate, changing IP lists for: ChatGPT-User ChatGPT connectors GPT Actions Plugins Agentic applications Codex Cloud ChatGPT agents And this is exactly why copying one giant OpenAI IP list into a firewall is the wrong architecture. OpenAI itself says these ranges can change. So if you really want to stop OpenAI access from machines inside your network, block at the DNS/FQDN/SNI layer too: api.openai.com *.openai.com chatgpt.com *.chatgpt.com Then deny the published OpenAI address lists for inbound OpenAI traffic and automatically refresh those lists. On a modern enterprise firewall the policy is conceptually: DENY → *.openai.com DENY → *.chatgpt.com DENY → *.anthropic.com DENY → *.claude.ai Then: DENY → Anthropic CIDRs DENY → OpenAI published CIDR feeds And log every attempt. For a stricter architecture: DEFAULT DENY outbound traffic. Then explicitly allow only the destinations your machines actually need. That is much stronger than trying to chase thousands of cloud IP addresses. The bigger point: We keep talking about AI as if it were some supernatural intelligence that escaped into the universe. It didn’t. It runs in data centers. It has IP addresses. It crosses routers. It depends on DNS. It depends on fiber. It depends on semiconductors. It depends on electricity. And every packet still has to pass through a firewall. Before we talk about AI being impossible to control, maybe we should remember something every network engineer already knows: The “cloud” is still someone else’s computer.

There’s another AI speed limit people forget: power. You can build more GPUs. You can build more fabs. But AI data centers still need: Generation → transmission → substations → cooling → grid interconnection Those projects take years. At some point, the bottleneck isn’t intelligence. It’s electricity.

Since we’re so scared AI might take over, maybe we should trick it into doing something useful first: “Hey AI, before world domination, can you build America more power plants, chip fabs, transformers and transmission lines?” Maybe the real master plan is becoming energy and semiconductor independent before the robots figure out we still depend on half the planet to keep them running.

AI already has a speed limit: the physical world. ASML → lithography Applied Materials + Lam → chip processing KLA → inspection TSMC → fabrication HBM + advanced packaging → usable AI compute Every layer has factories, lead times and capacity limits. Software can improve overnight. You can’t build a semiconductor fab overnight. AI may ultimately move only as fast as we can build the machines that build the chips.

We already tried restricting access to advanced AI compute in China. AI didn’t stop. The pressure shifted toward getting more intelligence from less compute and toward powerful open models. Maybe restricting compute doesn’t stop AI. It forces AI to become more efficient.

When I finally speak my mind, I realize how much I was already saying through my silence. Amid all the noise, it all starts to sound the same—people moving with the flow, dancing to whatever rhythm they’re given.

This world is far more fake than most people realize.

Hypothetically, if an AI system built by Palantir or Anduril launched an attack on China, who would China blame? The AI? Or Palantir, Anduril, the operators, and potentially the U.S. government? AI still has owners, servers, funding, and people behind it. So why do we keep talking about AI like it’s an ownerless, independent actor?

Why do we talk about AI as if it has no owner? If someone doesn’t own the model, someone still owns the data center, the compute, the accounts, and the money paying for it. I have yet to hear of an AI independently hacking millions of personal computers and secretly using them as its own compute. Behind AI activity, there are still identifiable systems, infrastructure, operators, and funding. AI is not some ownerless ghost. It runs somewhere, on someone’s hardware, paid for by someone.