There you have it! In September, Anthropic's CEO Dario Amodei told the industry to slow down. Tod…
November 10, 2026 · 0 likes · 0 comments
Cybersecurity AI
There you have it! In September, Anthropic's CEO Dario Amodei told the industry to slow down. Today, Nov. 9, Anthropic shipped two new frontier models.
Full story, just the facts: https://lnkd.in/eaPcSip7
Claude Fable 5.2 and Claude Mythos 5.2. Anthropic says both generate output 30% faster than the 5.1 generation. Price drops 20%, from $10 to $8 per million input tokens, and from $50 to $40 per million output tokens.
Faster and cheaper. In the same breath as "slow down."
To be fair, the company says it was misread. "We called for pacing, not pausing," Dianne Penn, Anthropic's head of product management for research, told CNBC. Amodei's essay does say pacing "does not mean halting model training." It also says, in plain words: "We must slow the pace at which we improve the capabilities of AI models."
You decide how those fit together.
Here's the part that matters for security. Anthropic says the two are one model with different safeguards. Fable carries classifiers that block high-risk cyber work like exploit generation and binary vulnerability analysis. Mythos removes those restrictions, and goes only to vetted organizations in Anthropic's Cyber Verification and Life Sciences programs.
Same brain. Different guardrails. Credit where it's due, that's a serious attempt at control. But the whole thing rests on the classifiers holding and the vetting holding.
And the speed keeps going up. Remember METR, the outside evaluator I wrote about last week? On Nov. 5 it estimated Anthropic is improving its models on well-specified tasks about twice as fast as a year earlier, with caveats, and said Anthropic's redactions limited what it could report. Anthropic's own Nov. 2 amended IPO filing says it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI."
Anthropic says the models went through rigorous safety and alignment testing, with new evaluations and post-launch monitoring. That's their word. The lab is still largely grading its own homework, and the outside graders get the redacted version.
Meanwhile, every 20% price cut puts this capability in more hands. Defenders. And everyone else.
I build with these models every day. I love what they do. But oversight is moving at government speed and the models are moving at lab speed.
So plan for it. Assume the attacker on the other side has a model this good, this cheap, this fast. Shrink your patch window. Test your own systems with AI before someone else does.
Time to wake up!
Full story, just the facts: https://lnkd.in/eaPcSip7
Claude Fable 5.2 and Claude Mythos 5.2. Anthropic says both generate output 30% faster than the 5.1 generation. Price drops 20%, from $10 to $8 per million input tokens, and from $50 to $40 per million output tokens.
Faster and cheaper. In the same breath as "slow down."
To be fair, the company says it was misread. "We called for pacing, not pausing," Dianne Penn, Anthropic's head of product management for research, told CNBC. Amodei's essay does say pacing "does not mean halting model training." It also says, in plain words: "We must slow the pace at which we improve the capabilities of AI models."
You decide how those fit together.
Here's the part that matters for security. Anthropic says the two are one model with different safeguards. Fable carries classifiers that block high-risk cyber work like exploit generation and binary vulnerability analysis. Mythos removes those restrictions, and goes only to vetted organizations in Anthropic's Cyber Verification and Life Sciences programs.
Same brain. Different guardrails. Credit where it's due, that's a serious attempt at control. But the whole thing rests on the classifiers holding and the vetting holding.
And the speed keeps going up. Remember METR, the outside evaluator I wrote about last week? On Nov. 5 it estimated Anthropic is improving its models on well-specified tasks about twice as fast as a year earlier, with caveats, and said Anthropic's redactions limited what it could report. Anthropic's own Nov. 2 amended IPO filing says it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI."
Anthropic says the models went through rigorous safety and alignment testing, with new evaluations and post-launch monitoring. That's their word. The lab is still largely grading its own homework, and the outside graders get the redacted version.
Meanwhile, every 20% price cut puts this capability in more hands. Defenders. And everyone else.
I build with these models every day. I love what they do. But oversight is moving at government speed and the models are moving at lab speed.
So plan for it. Assume the attacker on the other side has a model this good, this cheap, this fast. Shrink your patch window. Test your own systems with AI before someone else does.
Time to wake up!