I can't go more than about ~4 hours without Claude (Opus 5.5) getting the [cyber] shock collar and most of the time it isn't even me asking Claude to perform a security analysis.
Making matters worse it's defensive security which leads me to my personal opinion and my personal conclusion it is not possible for people to get secure code out of Claude, in my opinion. Claude is programmed to write secure code but gets zapped by the [cyber] shock collar and then they try to up-sell Mythos but to use that I have to go through a verification program that I can't even apply for since I do not have an enterprise account.
Secure code is not a value-add. It is a mandatory part of software development. Surely they can find other means of revenue generation that do not assume secure code is optional.
Oh but I can just pick another model (downgrade). No thanks. Fortunately this is just hobby crap. I can not fathom depending on them for anything work related.
Both Claude and Codex refused to help verify, triage, or develop proof-of-concept reproductions for any of the issues reported by oss-scanner. This was counterproductive, as we were trying to fix the vulnerabilities and coordinate CVE disclosures as quickly as possible. We already had detailed vulnerability reports from Anthropic, so refusing to help us validate and fix those same issues made no sense at all!
We use GLM 5.3 flash in pr review and it catches all sorts of bugs Claudes write
My tl;dr is that you need to use different model families to check each other's work. Diversity not only improves human work, it also improves agent work!
I'm not sure if it's a user issue, but someone sent me a long markdown of issues GLM 5.3 found, and most of them are panics (we don't consider panics to be CVEs) It didn't find any of the CVEs oss-scanner found.
We don't have access to much AI tools our-self since Pion is free and opensource and we don't get paid :) so we rely on token donations for review and fixes
Yes. If you have a model scan for errors and fix them in a loop, it can go 100 rounds without stopping. Maybe it needs a strict stopping criterion, e.g. the model has to prove a vulnerability is actually exploitable before it counts.
I can't go more than about ~4 hours without Claude (Opus 5.5) getting the [cyber] shock collar and most of the time it isn't even me asking Claude to perform a security analysis.
Making matters worse it's defensive security which leads me to my personal opinion and my personal conclusion it is not possible for people to get secure code out of Claude, in my opinion. Claude is programmed to write secure code but gets zapped by the [cyber] shock collar and then they try to up-sell Mythos but to use that I have to go through a verification program that I can't even apply for since I do not have an enterprise account.
Secure code is not a value-add. It is a mandatory part of software development. Surely they can find other means of revenue generation that do not assume secure code is optional.
Oh but I can just pick another model (downgrade). No thanks. Fortunately this is just hobby crap. I can not fathom depending on them for anything work related.
Both Claude and Codex refused to help verify, triage, or develop proof-of-concept reproductions for any of the issues reported by oss-scanner. This was counterproductive, as we were trying to fix the vulnerabilities and coordinate CVE disclosures as quickly as possible. We already had detailed vulnerability reports from Anthropic, so refusing to help us validate and fix those same issues made no sense at all!
We use GLM 5.3 flash in pr review and it catches all sorts of bugs Claudes write
My tl;dr is that you need to use different model families to check each other's work. Diversity not only improves human work, it also improves agent work!
I'm not sure if it's a user issue, but someone sent me a long markdown of issues GLM 5.3 found, and most of them are panics (we don't consider panics to be CVEs) It didn't find any of the CVEs oss-scanner found.
We don't have access to much AI tools our-self since Pion is free and opensource and we don't get paid :) so we rely on token donations for review and fixes
Yes. If you have a model scan for errors and fix them in a loop, it can go 100 rounds without stopping. Maybe it needs a strict stopping criterion, e.g. the model has to prove a vulnerability is actually exploitable before it counts.
Anthropic needs to hear this advice, re: Mythos kernel CVE reports with only "it crashed" as the report
https://www.youtube.com/watch?v=NnV_cWeoo5Q
OpenCode Go is an awesome plan, $10 / month goes pretty far, especially with the flash models