I mean, there’s a good reason the first rules of firearm safety are to always treat a weapon as loaded, and to never direct the weapon at something you aren’t prepared to destroy. The key point being that you never know when some freak accident can happen with a loose pin, bad ammo, a broken spring, or just a person tripping and shaking the gun a bit too hard.
A gun should never go off by itself. You still treat it as if it can, because in the real world freak accidents happen.
Sure. The point is it’s entirely possible to use a firearm safely. There is no safe use for LLMs because they “make decisions”, for lack of a better phrase, for themselves, without any user input.
That is not at all how LLMs work. It’s the software written around LLMs that aide it in constructing and running commands and “making decisions”. That same software can also prompt the user to confirm if they should do something or sandbox the actions in some way.
Only if the user has configured it to bypass those authorizations.
With an agentic coding assistant, the LLM does not decide when it does and doesn’t prompt for authorization to proceed. The surrounding software is the one that makes that call, which is a normal program with hard guardrails in place. The only way to bypass the authorization prompts is to configure that software to bypass them. Many do allow that option, but of course you should only do so when operating in a sandbox.
The person in this article was a moron, that’s all there is to it. They ran the LLM on their live system, with no sandbox, went out of their way to remove all guardrails, and had no backup. The fallout is 100% on them.
if you’re denying access to your agentic AI, what is the point of it? It needs access to complete agentic tasks.
Yes, which it can prompt you for. Three options:
Deny everything
Prompt for approval when it needs to run a command or write a file
Allow everything
Obviously optional 1 is useless, but there’s nothing wrong with choosing option 2, or even option 3 if you run it in a sandbox where it can’t do any real-world damage.
You can fine-grain nr. 2 even more: You can give access to e.g. modify files only in a certain sub-tree, or run only specific commands with only specific options.
A restrictive yet quite safe approach is to only permit e.g. git add, git commit, and only allow changes to files under the VC. That effectively prevents any irreversible damage, without requiring you to manually approve all the time.
They are not foolproof. They will absolutely cause problems in the hands of a fool. But they will not cause problems all on their lonesome. They’re inanimate objects. They cannot do absolutely anything without interaction from the user. If you can’t understand this, you should never be allowed to own one.
And neither can anthropic claude. Claude isn’t randomly deleting people’s websites, the kid gave anthropic bad instructions, it didn’t spontaneously decide anything. This is like an idiot pointing a gun at something he didn’t want destroyed and sneezing causing a trigger squeeze and then trying to blame the gun manufacturer.
They absolutely do not.
I mean, there’s a good reason the first rules of firearm safety are to always treat a weapon as loaded, and to never direct the weapon at something you aren’t prepared to destroy. The key point being that you never know when some freak accident can happen with a loose pin, bad ammo, a broken spring, or just a person tripping and shaking the gun a bit too hard.
A gun should never go off by itself. You still treat it as if it can, because in the real world freak accidents happen.
Sure. The point is it’s entirely possible to use a firearm safely. There is no safe use for LLMs because they “make decisions”, for lack of a better phrase, for themselves, without any user input.
That is not at all how LLMs work. It’s the software written around LLMs that aide it in constructing and running commands and “making decisions”. That same software can also prompt the user to confirm if they should do something or sandbox the actions in some way.
It can, but we’ve already seen many times that it does not.
Only if the user has configured it to bypass those authorizations.
With an agentic coding assistant, the LLM does not decide when it does and doesn’t prompt for authorization to proceed. The surrounding software is the one that makes that call, which is a normal program with hard guardrails in place. The only way to bypass the authorization prompts is to configure that software to bypass them. Many do allow that option, but of course you should only do so when operating in a sandbox.
The person in this article was a moron, that’s all there is to it. They ran the LLM on their live system, with no sandbox, went out of their way to remove all guardrails, and had no backup. The fallout is 100% on them.
As I said elsewhere, if you’re denying access to your agentic AI, what is the point of it? It needs access to complete agentic tasks.
No disagreement there.
Yes, which it can prompt you for. Three options:
Obviously optional 1 is useless, but there’s nothing wrong with choosing option 2, or even option 3 if you run it in a sandbox where it can’t do any real-world damage.
You can fine-grain nr. 2 even more: You can give access to e.g. modify files only in a certain sub-tree, or run only specific commands with only specific options.
A restrictive yet quite safe approach is to only permit e.g.
git add,git commit, and only allow changes to files under the VC. That effectively prevents any irreversible damage, without requiring you to manually approve all the time.And then when you give it access, it fucks shit up. I don’t know why this is hard to understand.
“Guns are foolproof”
You should have yours taken away.
They are not foolproof. They will absolutely cause problems in the hands of a fool. But they will not cause problems all on their lonesome. They’re inanimate objects. They cannot do absolutely anything without interaction from the user. If you can’t understand this, you should never be allowed to own one.
And neither can anthropic claude. Claude isn’t randomly deleting people’s websites, the kid gave anthropic bad instructions, it didn’t spontaneously decide anything. This is like an idiot pointing a gun at something he didn’t want destroyed and sneezing causing a trigger squeeze and then trying to blame the gun manufacturer.
LOL and you know this how?
No, this is more like pointing a gun downrange and then the gun fires itself and the bullet decides to do a U-turn and shoots the user.
Not really.
If you have the agent installed, it’s like having your gun assembled.
If you have your agent enabled, it’s like having your gun loaded.
If you give your agent permissions, it’s like taking your gun off safety.
If you don’t have your agent properly sandboxed, it’s like having bad muzzle control.
And if your agent is actively running, it’s like having your finger on the trigger.
This breaks every weapon safety rule. That’s how you get a negligent discharge.
Hence, it’s like scratching your back with a loaded weapon.