Elektrine lite

← Feed

@fiat_lux@lemmy.zip

Post #2835618

2026-05-25 13:33 UTC

In the latest episode of “behold the power of Mythos” from The Hacker News - Claude Mythos AI Finds 10,000 High-Severity Flaws in Widely Used Software I distilled it so you don’t have to. Of these vulnerabilities, 6,202 have been classified as high- or critical-severity flaws impacting more than 1,000 open-source projects. That 10,000 count didn’t even survive until paragraph 3. Subsequent analysis of these [6202] vulnerability candidates has identified that 1,726 are valid true positives. Ah fuck. 1726. But wait, a bad infographic has entered the ring! 23,019 potential vulnerability candidates Ok now we’re talking. 1,900 Reviewed by external security firms Wait, what? Why those? Why only those? 1726 confirmed positive You couldn’t even cherry pick the valid ones? 467 reported to maintainers Where did the other 1259 go? Maybe this other part of the flowchart will go better… 1,129 reported direct to maintainers by Anthropic, at their request (May contain false positives) 1129 + 467 = 1596 total reported to maintainers Most of them just spammed at open source maintainers. Right. Maybe Anthropic’s media release has the goods! 1,752 of those high- or critical-rated vulnerabilities have now been carefully assessed by one of six independent security research firms, or in a small number of cases by ourselves Slightly higher than the 1726, and I’m guessing the internal reviews were from scanning Claude’s code, but ok, whatever. Of these, 90.6% (1,587) have proved to be valid true positives, and 62.4% (1,094) were confirmed as either high- or critical-severity 10% of 10000 is still something, right? On maintainers’ request, we sometimes disclose bugs directly, without further assessment. We’ve now reported 1,129 such unvetted bugs, of which Mythos Preview estimated that 175 were high- or critical-severity. I’m sure those maintainers enjoyed that 16% accuracy rate based on Mythos’ own estimations. But wasn’t that 1129 the bulk of your reports? We estimate that we’ve disclosed 530 high- or critical-severity bugs to maintainers so far. There are a further 827 confirmed vulnerabilities (estimated as high- or critical-severity in the same manner) that we’re aiming to disclose as quickly as possible. 530 is only a third of the reports you made to maintainers… 65 of those have been given public advisories The infographic says 88. I’d ask if they were massaging their financials like they massaged 65 advisories, but we know they are. 23,019 potential vulnerability candidates of all severities, 65 advisories. If you printed the code out and drunkenly threw darts at it you’d probably hit the same level of accuracy.

Replies (5)

  • @V0ldek@awful.systems 2026-05-25 14:23

    All that it tells me is that if you spent the same amount of resources on just fuzzing randomly picked OSS codebases you’d probably get better value for your buck.

    Open ##2836764

  • @froztbyte@awful.systems 2026-05-25 14:37

    it continues to be amazing to me that this is the “high impact” area they’re going with: even if their analysis systems are better (and frankly I still don’t buy this wholesale, there’s a whole rest of the owl being handwaved[0]), but-elimination is by definition diminishing returns so you can only fanfare like this the first time [0] - having fucking gigantic budgets to throw at running a parse of every single repo and every test condition/simulation you wish to certainly does help a hell of a lot, even moreso when you can shell out to a half-dozen second stage review corps…

    Open ##2837037

  • @CinnasVerses@awful.systems 2026-05-25 15:04

    Anthropic (who own Claude Code) are hoping to IPO this year.

    Open ##2837656

  • So what’s the over/under on the discrepancies between the numbers that the HN folks got and the official press release numbers being in part due to some kind of hallucinatron hijinks? Because I’m gonna go ahead and predict with confidence that either the HN post was written with a faulty slopbot and they didn’t check it or else the presser itself went through the matrix-multiplication-meaning-mangler. Possibly both and all those numbers are similar levels of “more or less right, we swear”

    Open ##2859061

  • @schnoopy@awful.systems 2026-05-26 07:19

    1 cve, 100 things that might have mattered. 2 orders of magnitude false positives doesn’t sound like an efficient use of labour for finding vulnerabilities but that’s just me.

    Open ##2859454