sure they are doing """alignment""" to the models, and maybe they have some more sophisticated serverside mitigations. but the fact that the system prompt text is in the package at all rather than all being entirely serverside does the opposite of inspire confidence. Even the system prompt is fine with hacking as long as you go "it's ok I am good"
https://neuromatch.social/@jonny/116325221458366596

so this simultaneously raises the floor of doing open source at all to "if you can afford brute force generating exploits against your repos for days at a time" while simultaneously causing so many false positives that bug bounties are crumbling and the info giants will pull labor from open source projects by just generating them badly in-house - don't roll your own crypto becomes "now you have to roll your own crypto because nobody else is, and then pay an AI company to secure it for you."

you know that problem where it's actually in Google's best interests to sabotage their traditional search results to force everyone to use the AI results because then you never leave the site and direct prompt advertising becomes extremely valuable? yeah, it's like that for code, where it's actually in anthropic's best interests for all the code to be entirely unmaintainable and unsecurable except for with LLMs