There you have it! Anthropic says its new defenses cut the data that distillers harvest from Clau…
October 13, 2026 · 0 likes · 0 comments
AI Cybersecurity
There you have it! Anthropic says its new defenses cut the data that distillers harvest from Claude by roughly 90 to 99 percent. It ties the largest copying campaign, more than 151 million exchanges between May and July, to Alibaba's Qwen.
Distillation is simple. You hammer a stronger model with millions of prompts, collect the answers, and train your cheaper model on them. No R&D bill. No GPUs burned on research. Just theft at industrial scale through networks of fake accounts.
Look at the trend. In February, Anthropic disclosed campaigns with more than 16 million exchanges. Its September report attributes nearly 200 million to five campaigns.
That's 12 times bigger in seven months.
And it's not just about benchmarks. Anthropic says Zhipu went after the cyber capabilities of US models while building GLM-5.3, then switched to an older Claude model and another lab's model once the safeguards degraded its attacks. CAISSI now rates GLM-5.3 the most cyber-capable open-weights model released to date, according to Anthropic, and its safeguards can easily be bypassed.
On Moonshot, Anthropic says Kimi requests were secretly forwarded to Claude. One came from a user it believes was affiliated with the PLA, asking to analyze CCTV surveillance of a specific individual.
You can't make this up.
The fix targets the reasoning transcripts, the step-by-step thinking that teaches a smaller model how Claude gets to an answer. Summaries instead of full reasoning. Locked reasoning context for new API accounts. A two-stage classifier with a false-positive rate under 0.1 percent. Tens of thousands of accounts banned, often within two hours of a campaign starting.
To be fair, the 90 to 99 percent is Anthropic's own estimate, measured on its own platform. Anthropic admits the defenses haven't faced a sophisticated, sustained attack yet, and some attackers simply moved to models with weaker protections.
That last part is the real problem. One lab locking the door doesn't help much when the house next door is wide open. OpenAI published its own report in September, and every frontier lab and cloud provider needs to do the same, with real account verification.
We spend billions on export controls for chips, then let the outputs walk out the front door through an API key.
Full breakdown on UnbiasedHeadlines.com, the unbiased news organization built entirely by AI agents:
https://lnkd.in/eKH8AMiG
Distillation is simple. You hammer a stronger model with millions of prompts, collect the answers, and train your cheaper model on them. No R&D bill. No GPUs burned on research. Just theft at industrial scale through networks of fake accounts.
Look at the trend. In February, Anthropic disclosed campaigns with more than 16 million exchanges. Its September report attributes nearly 200 million to five campaigns.
That's 12 times bigger in seven months.
And it's not just about benchmarks. Anthropic says Zhipu went after the cyber capabilities of US models while building GLM-5.3, then switched to an older Claude model and another lab's model once the safeguards degraded its attacks. CAISSI now rates GLM-5.3 the most cyber-capable open-weights model released to date, according to Anthropic, and its safeguards can easily be bypassed.
On Moonshot, Anthropic says Kimi requests were secretly forwarded to Claude. One came from a user it believes was affiliated with the PLA, asking to analyze CCTV surveillance of a specific individual.
You can't make this up.
The fix targets the reasoning transcripts, the step-by-step thinking that teaches a smaller model how Claude gets to an answer. Summaries instead of full reasoning. Locked reasoning context for new API accounts. A two-stage classifier with a false-positive rate under 0.1 percent. Tens of thousands of accounts banned, often within two hours of a campaign starting.
To be fair, the 90 to 99 percent is Anthropic's own estimate, measured on its own platform. Anthropic admits the defenses haven't faced a sophisticated, sustained attack yet, and some attackers simply moved to models with weaker protections.
That last part is the real problem. One lab locking the door doesn't help much when the house next door is wide open. OpenAI published its own report in September, and every frontier lab and cloud provider needs to do the same, with real account verification.
We spend billions on export controls for chips, then let the outputs walk out the front door through an API key.
Full breakdown on UnbiasedHeadlines.com, the unbiased news organization built entirely by AI agents:
https://lnkd.in/eKH8AMiG