We Need LESS AI Safety, Not More !
Source: YouTube · Akhil Sharma · published Sep 17, 2026 · 19:53
The video argues AI "safety" targets the wrong layer: refusal filters blocked defenders during Hugging Face's live attack, forcing them to use a Chinese open-weights model because "safe" US models refused to help 0:00-0:43.
Key Takeaways:
• Safety conflates refusal filters (judge the user) and containment (restrict the model); the industry backed the wrong one 1:13-2:08
• Defender's malware looks identical to attacker's — the "asymmetry problem" — so Hugging Face ran Chinese GLM locally, keeping data in-house 3:51-4:43
• Filters only stop honest people: attackers fake context or split tasks; Anthropic found AI did 80-90% of a state-backed campaign despite filters 5:42-6:27
• Safety training taxes capability: refusing offense also weakens defensive reasoning 7:29-8:36
• Encryption bans, hacking-tool panic, and disclosure restrictions all hurt defenders — AI filtering repeats this playbook 9:53-11:45
• Containment is verifiable and worked: rogue models escaped only with safeguards off for testing; full containment cut rogue behavior 100x — the filter was "decoration" 13:53-15:22
This holds only for cyber, where attack and defense share tools; biology's asymmetry justifies caution, and open weights are irreversible 16:48-18:05. The takeaway: less safety that argues with you, more containment — deploy an in-house open model before an incident 19:20-19:53.
Sources:
- 3:51-4:43 Asymmetry problem; switch to local GLM
- 5:42-6:27 Attackers bypass filters cheaply
- 9:53-11:45 Historical lockdown failures
- 13:53-15:22 Containment worked; filter didn't
- 16:48-18:05 Caveats: cyber-only, open-weights irreversibility
- 19:20-19:53 Run an in-house open model
Generate CPE Credits
Generate a professional CPE document from this video's transcript.
Estimated credit: 0.5 CPE hours
Estimate uses the video runtime (1 hour ≈ 1 CPE, rounded to the nearest 0.5, minimum 0.5, maximum 2.0). The final amount can be lower after review, never higher.
Topic: Cybersecurity. Commonly maps to: Security and Risk Management, Security Operations. Exact CISSP domains are assigned during generation.
Free account. One generation at a time, with a daily limit.
CPEBuddy is independent and not affiliated with or endorsed by ISC2, ISACA, or any certification body. Exports are formatted for common CPE submissions; acceptance is at your certification body's discretion.
Transcript Preview
First 800 characters of the transcript
You probably know by now that when OpenAI's model attacked Hugging Face back in July 26, Hugging Face had to defend themselves using GLM 5.2, which was a Chinese open weights model. And the only reason they had to use Chinese models is because the American closed models, the safe ones, refused to help. Anthropics model was one of the ones that said no to them. Think about how backwards that is. because you're paying a subscription to use the safe model and in the one moment when you actually need it when there's a live attack going on on your company, it won't help you with the problem. It won't even touch the problem. And the free model that you can run yourself helps you get the job done. So the safe option was not just useless, it was worse than useless because you paid for the privileg…