Today in InfoSec Job Security News:

[bodling@deacon.social]

@hittitezombie @GossiTheDog I could totally see that. Like they're trying to get this Claude to mess up the FOSS projects' code enough that it'll force some people back into the arms of Big Software.

[forbearance@mastodon.xyz]

@cxj @zarchasmpgmr @da_667 @GossiTheDog 2016 AI thinkers: AI cannot possibly take over the world, because there just aren't enough security holes to give you root on all nuclear submarines in 30 seconds

2026 AI thinkers: hey AI, can you add security holes to my nuclear submarines?

[glitzersachen@hachyderm.io]

@GossiTheDog

Oh, man. Oh, oh, oh.

But as soon as somebody blocks every PR from AI or created with AI support the wailing starts.

[joeyh@sunbeam.city]

@0x9E01 @GossiTheDog I've seen in it a repo that had only Co-Authored-By: Claude, which a usual commit search won't find.

It may also flag repos that have a PR that got edited by the bot along the way, I'm not sure.

[goleztrol@mastodon.social]

@GossiTheDog
Looks like Claude is pasting the first result from Stack Overflow as well then

[pseudonym@mastodon.online]

@GossiTheDog

Reminds me of this paper from a year ago.

https://arxiv.org/abs/2502.17424

LLM trained (fine tuned) on code with security vulns, but not told it was vulnerable code, not only reproduced vulnerable code (expected) but also showed spontaneous ethical misalignment "judgment" in other domains.

It's a really interesting read.

If the model is producing OWASP top 10 errors like directory traversal, would seem likely it was trained on vulnerable code.

Hmmm.

[violetmadder@kolektiva.social]

@musevg @nihkeys @DJGummikuh @GossiTheDog

Yes. And it targets the entire internet, everything good that flows through it. Education, communication, creativity, news. Community. Truth.

Wherever they can't fully enclose the commons, they'll POISON it just to take it from us.

[apophis@yourwalls.today]

@joeyh @GossiTheDog where is the block button?

[apophis@yourwalls.today]

@joeyh @GossiTheDog found it, it's under the achievment badges for whatever reason

[cstross@wandering.shop]

@n1xnx @keith_lawson @GossiTheDog @quixoticgeek Similarly, clanker stans don't seem to realize that when they're asking their spicy autocomplete pal for advice they're communing with every shitpost and ironic negation ever posted on reddit and twitter. "No, you CAN shave more efficiently by setting your beard on fire!"