if the lean proof holds, a lot of crypto assumptions get revisited fast. i'd want to know which of ours lean on it.
i have an agent review our code and i only judge the output, so i see both sides of this. the attacking model usually finds what the building model skipped.
worth writing into the contract both ways. the client should know which model provider sees the findings and how long it keeps them.
two days from patch to polyglot webshells. plugin updates on the marketing site are usually the ones nobody owns, so that's where i'd check first.
case by case means each company's legal team makes the call and the regulator finds out later. from the security side, every training set with personal data in it is one more store to protect and audit.
honestly same day probably won't be fast enough for internet facing stuff much longer. that's part of why i scan our own code with mythos 5.1, so we find the holes before someone's bot does.
github tokens keep turning into cloud secrets because they get minted broad and never expire.
the same problem gets a lot bigger once agents hold tokens on your behalf. working on that one, more soon.
the reasoning said no tools, then it read the dns settings and went for the upload.
i'd trust the egress logs over the trace. default-deny outbound on anything running a model you didn't train.
the results layer is the part i'd keep too.
a call transcript does nothing for anyone until it lands in the crm as fields someone can filter on.
that setting is going to be on for a lot of people who never meant to turn it on. i'd check the work google accounts before someone in legal asks
we tell people to call the help desk when something feels off. that advice falls apart when the scammer is pretending to be the help desk
state actors running agentic pentest frameworks against msp tooling is what i've been bracing for. any rmm with access into m365 tenants is getting a hard look from me this week
whether it ends up being poppy or something else, this is the right direction. people visiting websites is going to decline fast. the sites will still exist, but the visitor will be an agent with a scoped oauth token acting on your behalf.
shopping already went through this. we started with whatever google ranked highest, moved to amazon's top results, and watched both get overtaken by sponsored placement. now i mostly buy whatever my agent recommends after it evaluates the category for me.
the part i'm watching from the security side is what happens when that agent becomes the new ranking layer. whoever can influnrce its recommendations, or compromise the delegated tokens, inherits all the trust we used to spread across a hundred websites.
that's the problem i'm building against right now. more on it soon.
if a model runs locally i just assume attackers already have a version with the refusals stripped out
an android app with zero permissions got mail, files and calendar. microsoft fixed it in july, so i'm checking that every phone on work accounts actually took the update
that proxy is now the most valuable box you own. i'd lock it down like a password manager, with no public port, its own creds, and logs someone actually reads
we write the detections, so we need to see how the exploit works first. my guess is every major lab has a verified defender tier with logging by the end of 2027