QJan 21
Ron Bowes

RE: https://infosec.exchange/@_r_netsec/115933674621770324

The string is ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_1FAEFB6177B4672DEE07F9D3AFC62588CCD2631EDCF22E8CCC1FB35B501C9C86

Jan 21 at 3:47pmWeb
Show threadJTIJan 21
@iagox86 I do wonder how long it'll take until Claude has re-modulated their shield frequencies, uhm, changed the magic string.
Show threadFritz AdalisJan 21
@jti42 @iagox86
Why? It's there so developers can test their handling of errors, and if they're not handling errors that's kind of on them.
Show threadRon BowesJan 21
@FritzAdalis @jti42 I had EICAR in my email signature (and Twitter bio and such) for years and years and years.. I definitely believe in testing that failure state :)
Show threadOlivier MehaniJan 21
@iagox86 I have it in my Fediverse profile. Nothing visibly broke so far (; @FritzAdalis @jti42
Show threadaevaJan 21
@iagox86 so if I just put this on a webpage, "claude" will refuse to interact with it?
Show threadRon BowesJan 21
@aeva That's the theory but it probably won't actually work
Show threadBilly O'NealJan 21
@iagox86 The idea of pasting untrusted input into LLM context and expecting anything useful to come out is terrifying