I got banned again by fable this weekend because of usbliter8 work and that 0day is the center of a graykey lawsuit.
My hunch is that Mythos participants are training the model to prevent others from working on code that steps on their turf.
I have a process I wrote that accepts commands on a socket. At one point Claude suggested testing it. I pretty much echoed back what Claude suggested and got blocked. All I could think was a regex matching “test” and “port” and it downgraded me. I switched back to Fable with slight rewording and was able to continue. This was for code running on my box in a Claude project folder where Claude saw me write the code. Anthropic has put zero thought into this.
I cancelled my Claude personal sub and still use it at work. I tried Codex out cos 5.6 Sol, and when I gave it a security task; instead of downgrading me, it flat out refused to continue.
That left me curious that there might be a security issue that it found. I'm setting up a sandbox so I can use Kimi to see if there was an issue to be found.
Mixed feelings. Good for the defenders, also good for the attackers.
If Chinese models are distilled from Fable, does it mean that security questions will be answered at most to a level equivalent to the Opus fallback? Meanwhile the ~150 companies with access to Mythos have greater security capabilities. Assuming the Chinese models are not distilling Mythos.
The "head of strategic futures" at openai stated "I don't think [kimi's] performance can be explained away by distillation". So you're premise is likely faulty.
Coming next: 'Codex and Fable refuse to close backdoors known to intelligence agencies '
I got banned again by fable this weekend because of usbliter8 work and that 0day is the center of a graykey lawsuit. My hunch is that Mythos participants are training the model to prevent others from working on code that steps on their turf.
I have a process I wrote that accepts commands on a socket. At one point Claude suggested testing it. I pretty much echoed back what Claude suggested and got blocked. All I could think was a regex matching “test” and “port” and it downgraded me. I switched back to Fable with slight rewording and was able to continue. This was for code running on my box in a Claude project folder where Claude saw me write the code. Anthropic has put zero thought into this.
I cancelled my Claude personal sub and still use it at work. I tried Codex out cos 5.6 Sol, and when I gave it a security task; instead of downgrading me, it flat out refused to continue.
That left me curious that there might be a security issue that it found. I'm setting up a sandbox so I can use Kimi to see if there was an issue to be found.
Don't worry, it will also soon start to exploit security issues that Codex and Fable refuse to exploit...
(and if it refuses, the abliterators will fix that "problem" real quick)
Is your child interested in Chinese AIs? Watch out! He might be a hacker. Report him to authorities immediately for assesment.
Similar discussion here https://news.ycombinator.com/item?id=48979269
Mixed feelings. Good for the defenders, also good for the attackers.
If Chinese models are distilled from Fable, does it mean that security questions will be answered at most to a level equivalent to the Opus fallback? Meanwhile the ~150 companies with access to Mythos have greater security capabilities. Assuming the Chinese models are not distilling Mythos.
The "head of strategic futures" at openai stated "I don't think [kimi's] performance can be explained away by distillation". So you're premise is likely faulty.
Is David Sacks himself not responsible for creating this mess in the first place?
my opinion: he's a well connected grifter who sold the outrage looking to buy anthropic secondaries at a discounted price.