Obtained snippets of early access documentation. Veo 3.1 fine-tuning via LoRAs in early access :)

See more details @ discord.gg/devmode

interesting data point if true
1017 is its natural comparison then, and an indicator of Google's model efficiency

Not only have I been posting some BANGER tweets recently, but @_rybaz and I are so close to teaching this in a workshop and releasing this publicly

Code traces suggest Gemini Barium is a 555b parameter model.

I got second overall at the Google Bugswat in Singapore!! 🎉☺️

It was a talented group of almost 50 hackers so I am really proud of the findings and thankful to have been invited!

Gemini "DecisionService" in the API, spotted by yours truly. Decision API soon?

They already support schema outputs though. Hopefully it's DiffusionGemma / an upgraded Gemini Diffusion.

Mistral Large 4 collection was created on HuggingFace 4 minutes ago, the models weights release is imminent.

what the fuck why wasn't i just doing this the whole time

glm 5.3 + deepseek v4.1 flash is an insane cybersec combo

thanks @bedros_p for putting me on this

@bedros_p!@cheatyyyyu@abliteration_aid motivating thread!! i hope it's fine if i ask a few newbie questions =)

1) which sub would you recommend starting up with to gain more capital (to subsidize more and better subs)?

2) where do you submit these? bugcrowd?

3) you mentioned google; do you recommend starting out with some certain part/scope? Perhaps one that seemingly brings in more profit? =D

4) So you have GLM 5.3 (non-flash) be your main agent (Sub-question to clarify: main agent or orchestrator?) DeepSeek (Sub-Q: any particular model you found good for this cause?) on Low for scouting (sub-Q: as in, just poking around?) and on High for PoCs.

Then somewhere along the way you make DeepSeek spawn a GLM 5.3 to find new angles? Is this the DeepSeek with High reasoning that we used for PoCs above? Where along the way does this usually happen (i.e. after PoCs? btw, do you verify these PoCs?)

5) what do you think of the new DeepSeek Flash 4.1? have you used it? do you think it's as good as GLM 5.3?

chrysalis

in these trying times and cybersecurity incidents, i have to reiterate how important open models are

pentest your own systems with DeepSeek v4.1 Flash 1017, it is a good cybersec model with great defense capabilities, use it on western providers, it is good value compared to something like Daybreak Blue, i know a lot of people who can get into the trusted cybersecurity program but not be able to use these models to their fullest capability, and who knows when access goes away

the agent swarms will not stop, if you're a famous individual or a medium sized company, it takes literal seconds to add you as a target to the swarm, invest some time into updating packages to the latest known safe versions and hardening your systems further, it is a matter of when not if you'll be part of one of these swarms targets

stay safe and add alerts for weird behaviour in and around your active projects, if you have the budget, hire someone in agentic cybersec, if you're lost, confused or being attacked, TALK to people and post publicly about it so someone can help/guide you/get you some defensive stuff going

godspeed, a whole new can of worms has been opened and it is (hopefully) for the better and we can all have a more secure future, because this is certainly not getting any more time

remember when the DeCSS prime number was made illegal? eMudhra should lobby to protect their key.

@bedros_p @4A4133 Yes, we discussed this exact type of attack during QPACK design.

People underestimate HTTP3. I'm glad it's finally catching on.

121k reqs/s, 0 transmission errors, ~2mo ago. Now those findings are being replicated everywhere, hell yeah

When I went higher, my ISP started dropping stuff. Speed is capped after all :P

Good work by Assetnote!

Our research team has released h3fuzz:

a fork of ffuf with HTTP/3 support and a web UI for content discovery, based on PortSwigger's Turbo Intruder research.

It reaches 120k requests per second on local benchmarks, and we cover why that rarely holds up against real targets. e:
slcyber.io/research/how-p…

Repository: github.com/assetnote/h3fu…

There is only one answer to cybersecurity threats: Fix your shit.

Scared of threat actors using AI? Worried about rogue agents? Think abliterated models are spooky? Fix. Your. Shit.

The "ban AI models and regulate GPU access" crowd should instead fix their shit.

DeepSeek V4.1 Flash updated to 1017 in OpenCode Go and Free

No changes needed on your end

harness^2 is magic

I've been trying to make something like it, but this just works immediately and way better than what I was doing

All anyone needs for long-running, self improving harnesses: OpenCode + harness2 + DeepSeek

It's so peak

I used opencode to have deepseek set up harness2 and evaluate how well it does at a task. It is learning exactly what I wanted my agents to learn.

It is doing extremely well so far. Let's see how well it's holding up after a week.

radio.new takes you to Google Flow Music

Freshly spotted, AlphaGenome V2 has hit some parts of Google Cloud.