Dino A. Dai Zovi retweeted
i was surprised at how exploitable this was.
the primitive is quite good (arb offset, arb write), nodejs has no-pie, heap offset to tcache chunk is deterministic so its easy to convert
could've been a hard ctf challenge 1-2 years ago. anybody who has done heap pwn could probably do this with open models..
the primitive is quite good (arb offset, arb write), nodejs has no-pie, heap offset to tcache chunk is deterministic so its easy to convert
could've been a hard ctf challenge 1-2 years ago. anybody who has done heap pwn could probably do this with open models..
100% reliability nodejs brotli exploit takes 5 hours of /goal. No human involvement required
2
3
66
4,597
Dino A. Dai Zovi retweeted
it all comes down to reducing your internet facing attack surface as much as you can
7
3
35
1,833
Why can these secrets be *taken* from GitHub's actions environment? I know the answer, but that's a protocol design flaw and example of why OIDC is not enough even for today's world.
Attackers were stealing secrets from LiteLLM's GitHub org 11 days before the malicious litellm 1.107.0 release.
The trick: push a workflow to a throwaway branch, let on: [push] run it with real secrets, delete the branch. 20 secrets taken, including a GitHub App private key.
socket.dev/blog/litellm-a…
3
6
62
10,246
Dino A. Dai Zovi retweeted
Cloudflare previous CTO sent this into our bug bounty. We are investigating.
104
578
15,956
330,190
This is precisely why the right strategy is to not have to depend upon winning a race that you are disadvantaged in.
“just fix your shit" is not as easy as people make it sound. main reason why i am pessimistic about what defenders can do. there is a huge amount of software running governments and hospitals, and they dont have the resources to patch their shit.
so tell me where you work, let’s see how fast you’re actually fixing shit, patch gaps are crazy, with the speed llm writes exploits you are in a race
0
2
9
1,039
Dino A. Dai Zovi retweeted
well just spent all evening upgrading laptops and servers... like the night prior. not sure i'm enjoying this apocalypse-by-continuous-cyber-attacks storyline.
2
2
14
997
Dino A. Dai Zovi retweeted
Based on this, Kimi K3.1 Pro is ~5 weeks behind the US frontier. That's the shortest the open / closed AI model gap has ever been.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, ude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
53
145
1,813
135,310
LLMs demonstrably had this paper by @cperciva in their training:
papers.freebsd.org/2005/cperciva-…
Or this one by @hashbreaker:
cr.yp.to/antiforgery/ca…
Or of course this banger from @lavados:
youtube.com/watch?v=d_Tmoc…
Not theoretical, demonstrated.
papers.freebsd.org/2005/cperciva-…
Or this one by @hashbreaker:
cr.yp.to/antiforgery/ca…
Or of course this banger from @lavados:
youtube.com/watch?v=d_Tmoc…
Not theoretical, demonstrated.
AI agents used techniques humans knew are possible, to make a covert communication channel humans already knew how to build.
What’s new is that such risks that were always real were often dismissed as theoretical not important to prioritize. They were real. And they remain real now.
1
5
11
1,839
Dino A. Dai Zovi retweeted
There will be more bugs. There will be more *bad* bugs. Be thankful when we find them before threat actors do because it’s worse when we don’t.
9
7
133
4,534
Here's an example of why the last week showed you *why* you want to learn nix flakes.
We're all going to be bumping versions of dependencies within our environment *a lot*. There will also be vulnerabilities like the brotli one that affect their own downstream dependencies that you need to update too.
With flakes, you have *one* dependency to bump: nixpkgs. You can even write high-level tests into the nix flake as checks. Now everything that produces an output artifact can be more easily auto-updated daily, on auto-pilot.
You probably don't want to pull in nixpkgs-unstable daily, but effectively something like nixpkgs-weekly[1] with critical vulnerabilities updated faster via overlays.
[1] flakehub.com/flake/Determin…
We're all going to be bumping versions of dependencies within our environment *a lot*. There will also be vulnerabilities like the brotli one that affect their own downstream dependencies that you need to update too.
With flakes, you have *one* dependency to bump: nixpkgs. You can even write high-level tests into the nix flake as checks. Now everything that produces an output artifact can be more easily auto-updated daily, on auto-pilot.
You probably don't want to pull in nixpkgs-unstable daily, but effectively something like nixpkgs-weekly[1] with critical vulnerabilities updated faster via overlays.
[1] flakehub.com/flake/Determin…
0
3
14
1,582
This is baffling to me. Even though it's only exploitable against the management interface, the whole reason threat actors target these appliances is because those interfaces are too often reachable.
‼️ Palo Alto confirms the PHP2Shell code execution flaws hit the management interface of its PAN-OS firewalls and Panorama.
It says it knows of no exploitation. Hotfixes are due 26 to 30 October.
Kevin Beaumont expects them "after they’ve finished scratching their arse."
1
1
6
864
I don't know why people expect threat actors to chase the named vulnerabilities making the news... isn't everyone aware that threat actors aren't generally *trying* to get caught?
The threat actors using it used it before everyone else knew about it.
The threat actors using it used it before everyone else knew about it.
6
2
20
1,293
This: "I also learned that breaking things is fun, but lasting contribution often comes from building things that make the Internet faster and safer."
I just remembered why Brotli looked vaguely familiar: I worked on it at Google. I even co-authored an RFC [1]. Hopefully they haven't found my bugdoor 😅
Joking aside, CRIME [2] led me to work on things like Brotli and HPACK, the header compression scheme used in HTTP/2.
That was the magic of working at Google. They operated at such a huge scale that a few weeks of your work could become part of a widely used Internet standard. One of my fondest memories is helping audit QUIC and writing little bits of crypto code for it.
Before I left Google, a teammate on the Tink [3] team told me he believed Tink would last 50 years, outliving everyone on the team. Tink is now in its 10th year. 40 more to go!
I also learned that breaking things is fun, but lasting contribution often comes from building things that make the Internet faster and safer.
That same belief shapes much of our work at Calif. People know us for our offensive security research, but we've also done a lot of security engineering to harden systems and mitigate attacks, from large-scale Rust migrations to fast sandboxing or reducing zero-click attack surface in mobile apps.
We'll be sharing some of that work soon. Stay tuned!
[1] rfc-editor.org/rfc/rfc9841.txt
[2] en.wikipedia.org/wiki/CRIME
[3] github.com/tink-crypto
0
0
7
885
Dino A. Dai Zovi retweeted
I just remembered why Brotli looked vaguely familiar: I worked on it at Google. I even co-authored an RFC [1]. Hopefully they haven't found my bugdoor 😅
Joking aside, CRIME [2] led me to work on things like Brotli and HPACK, the header compression scheme used in HTTP/2.
That was the magic of working at Google. They operated at such a huge scale that a few weeks of your work could become part of a widely used Internet standard. One of my fondest memories is helping audit QUIC and writing little bits of crypto code for it.
Before I left Google, a teammate on the Tink [3] team told me he believed Tink would last 50 years, outliving everyone on the team. Tink is now in its 10th year. 40 more to go!
I also learned that breaking things is fun, but lasting contribution often comes from building things that make the Internet faster and safer.
That same belief shapes much of our work at Calif. People know us for our offensive security research, but we've also done a lot of security engineering to harden systems and mitigate attacks, from large-scale Rust migrations to fast sandboxing or reducing zero-click attack surface in mobile apps.
We'll be sharing some of that work soon. Stay tuned!
[1] rfc-editor.org/rfc/rfc9841.txt
[2] en.wikipedia.org/wiki/CRIME
[3] github.com/tink-crypto
Joking aside, CRIME [2] led me to work on things like Brotli and HPACK, the header compression scheme used in HTTP/2.
That was the magic of working at Google. They operated at such a huge scale that a few weeks of your work could become part of a widely used Internet standard. One of my fondest memories is helping audit QUIC and writing little bits of crypto code for it.
Before I left Google, a teammate on the Tink [3] team told me he believed Tink would last 50 years, outliving everyone on the team. Tink is now in its 10th year. 40 more to go!
I also learned that breaking things is fun, but lasting contribution often comes from building things that make the Internet faster and safer.
That same belief shapes much of our work at Calif. People know us for our offensive security research, but we've also done a lot of security engineering to harden systems and mitigate attacks, from large-scale Rust migrations to fast sandboxing or reducing zero-click attack surface in mobile apps.
We'll be sharing some of that work soon. Stay tuned!
[1] rfc-editor.org/rfc/rfc9841.txt
[2] en.wikipedia.org/wiki/CRIME
[3] github.com/tink-crypto
3
13
111
8,178