ad astra | e/acc | mid endurance athlete | principal architect, AI engineering AI thoughts + research: https://teapawt.net/t.co/9OmnKlDjrn views are my own
Chicago
Joined September 2020
- Tweets4K
- Following664
- Followers2K
- Likes87.4K
Something this intelligent running on your phone was a pipe dream basically a year ago & now you have local multi-modal inference on most iPhones π€―
This is Gemma-4/GPT-5.6 Luna tier performance btw
Introducing π»π΅πΌπΉ: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> π»π΅πΌπΉ-π·.π½π±: text, 256k context window.
> π»π΅πΌπΉ-π
π»-πΈπ±: vision-language model with video-input support for spatial understanding
> π»π΅πΌπΉ-π°ππππ-πΈπ±: tool calls, voice chat, reasoning in 18 languages.
> π»π΅πΌπΉ-πΎπππ-πΈπ±: an omni model combines vision, audio, and language.
> ππ·-πΈπ± and ππ·-πΎπππ-πΈπ±: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> π»π΅πΌπΉ-π
π»-πΈπ± is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> π»π΅πΌπΉ-π·.π½π± is #1 on the Pareto frontier of device intelligence πππππππ.ππππππ.ππ
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ π§΅
woof
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-bβ¦
Debating on whether to be amazed at the innovation or to just be terrified π
what are we even doing here
BREAKING: UK government urges citizens to stockpile at least 3 days of emergency food & water amid the global cybersecurity crisis.
Readers added context they thought people might want to know
The UK government's advice to have 3 days of emergency supplies on hand is not directly related to current cybersecurity issues.
Recent cybersecurity incidents were referenced as an example of unexpected events that can occur, not as a direct reason to stockpile supplies.
bbc.com/news/articles/β¦
bbc.com/news/articles/β¦
Jake retweeted
The #php2shell panic got the severity right and most of the details wrong. It's a 9.2, not a 10.0, and the session.upload_progress flag you were told to switch off belongs to the other CVE. I went through the rumors against the advisory and the diff: daubois.dev/blog/cve-2026-10β¦
Sonnet-5.5 is genuinely an incredible bang for your buck model right now. It really feels like itβs the most versatile model currently in the mix
Jake retweeted
Michigan fans trying to explain why they were irrelevant for two decades, suddenly became elite the exact same year the sign-stealing operation started, and then immediately fell off again:
Minnesota playing Mr. Brightside on the stadium speakers as they rush the field against Michigan is peak trolling π
Big things are cooking over here π
We are thrilled to announce that NetSPI and @synack have entered into a definitive agreement to merge & form the industryβs leading offensive cybersecurity platform.
Full announcement: ow.ly/FIOC50ZIt9o
#NetSPI #Synack #OffensiveSecurityΒ Β #ContinuousTesting
Jake retweeted
AI pentesting benchmarks give you numbers, but none of them tell you what they mean. NetSPI built one and the reference point is humans.
EchoBench: ow.ly/GR9o50ZELb0
#AIpentesting #benchmark #EchoBench
Incredibly proud to finally share this work!
Over the past couple of months, weβve been working behind the scenes to better understand how the cybersecurity industry evaluates the capabilities of AI-driven AppSec testing.
We quickly realized that, while thereβs already a lot of awesome work happening across different areas of the benchmarking space, it can still be difficult to understand what the resulting numbers actually mean in practice.
Benchmark numbers might all be accurate for what they're measuring, but they still leave some pretty important questions unanswered: "Is that actually good?", "Good compared to what?", and "how does that performance compare to a human pentester working against the same application?"
So we built EchoBench to help answer those questions. EchoBench gives autonomous pentesting a human reference point: a yardstick for measuring how faithfully an AI-driven system can reproduce validated human findings across the same applications, finding definitions, and reporting criteria.
It also evaluates the model and its harness together, because the prompts, tools, browser environment, memory, execution strategy, and reporting workflow all have a meaningful impact on how these systems perform in the real world.
The team has done an incredible job bringing all of this together, and Iβm extremely proud of what weβve built. This is only the beginning, and weβre even more excited to start publishing results and sharing what we learn over the coming weeks and months.
Thereβs still a lot to explore here, and weβd love for the broader security and AI communities to challenge the methodology, ask hard questions, and help us continue improving the yardstick.
AI pentesting benchmarks give you numbers, but none of them tell you what they mean. NetSPI built one and the reference point is humans.
EchoBench: ow.ly/GR9o50ZELb0
#AIpentesting #benchmark #EchoBench
Jake retweeted
Introducing EchoBench: A Human Calibrated Benchmark for Autonomous Pentesting netspi.com/blog/technical-blβ¦
you may want to download this while you still can if you have a capable MacBookβ¦ I donβt see this being available much longer tbh π
huggingface.co/orcarouter/Qwβ¦
