execveat

Defenders have threat modeling on their side. With access to source code and design docs, configs, infra, actual requirements and ability to redesign / choose the architecture and dependencies for the job, etc - there's a lot that actually gives defending side an advantage.

I'm quite optimistic about AI ultimately making systems more secure and well protected, shifting the overall balance towards the defenders.

The real world use cases for LLM poisoning is to attack places where those models are used via API on the backend, for data classification and fuzzy logic tasks (like a security incident prioritization in a SOC environment). There are no thumbs down buttons in the API and usually there's the opposite – promise of not using the customer data for training purposes.

CrowdStrike does this trick where it replaces the file (being transferred over a network socket) with zeroes if it matches the malware signature. Assuming that these are the malware signature files themselves, a match wouldn't be surprising.

It's heartwarming to believe in the "family" narrative some companies promote, but it's important to remain pragmatic. This narrative is just a motivational tool to encourage employees to go above and beyond their compensated duties. It's not a binding contract though.

In reality, the power dynamic inherently favors the employer. Once an employee has invested their time and energy, the company holds all the leverage. There's little incentive for the company to uphold their end of this unspoken "deal."

Leadership changes, company priorities shift, and the "family" narrative can quickly fade when faced with financial realities or strategic decisions.

I might be misunderstanding, but from Andrew's Linkedin it looks like he wasn't a security researcher at MS, he was actually the person responsible for translating between security researchers and the upper management:

Evangelize security services, practices, products, both internally and externally.

Leading technical conversations around strategy, policy and processes with FINSEC and DoD/IC executive staff.

For the MS size entities, the risk calculation is way more complicated. The 1:1 between cost of mitigation vs cost of exploitation only applies to opportunistic attacks, really. At the level where APTs get involved, the data / access might be so valuable that they'd gladly outspend blue team's budget by a factor of 10-100.

Well, because as a security person I can only evaluate his actions from the point of security. Evaluating actions of MS business leadership is beyond my expertise.

I highly doubt that the senior leadership would willingly accept this kind of liability. But you need to put it into right terms for them to understand. Politics play important role at that level as well. There are ways of putting additional pressure on the c-suite, such as making sure certain keywords are used in writing, triggering input from legal or forcing stakeholders to formally sign off on a presented risk.

Without insight knowledge, it's impossible to figure out what went wrong here, so I'm not assigning blame to the whistleblower, just commenting that way too often techies fail to communicate risks effectively.

I work in infosec, and this sounds like a communication failure on the whistleblower's part.

Contrary to what many people believe, the profits should be prioritized over security for the most companies, that's only natural (after all, they don't generate any profits themselves, typically). The key is finding the right balance for this tradeoff.

Business leaders are the ones that are responsible for figuring out the acceptable risk level. They already deal with that every day, so it's nonsensical to claim they aren't capable of understanding risk. InfoSec's role for the most part is being a good translator, by identifying the technical issues (vulnerabilities, threats, missing best practices) that go beyond the acceptable risk profile and to present these findings to the business stakeholders, using the language they understand.

Either the guy wasn't convincing enough, or he failed to figure out the things business cares about & present the identified risk in these terms.

At this point you should just leave this dumpster fire of an organization and find a more reasonable place to work. I can't relate to the people who keep inventing atrocious workarounds ignoring the problem that they work in a hostile work environment.

I work in security and can't relate to banning Python & replacing it with Microsoft crap either.

Allow the eIDAS certificates, but limit them to the country code TLDs to match the jurisdiction of the certificate issuer.

You can't block a browser at the DNS level.

No, there is no confusion here at all (for a Python developer). I would consider it a code smell though as the whole problem is completely avoidable by better naming.

By the way, there is a vulnerability (Prototype Pollution) that is only possible due to this behaviour in JS: https://portswigger.net/web-security/prototype-pollution

There also are programmers who are aware of much more complicated things being done in sister projects, like JVM and JS runtimes in this case.

In my experience Bing chat and phind are useless. But perplexity.ai and GPT4 are amazing. GPT-3.5 and Cloude-instant (available through poe.com) are cool as well, even though they got significantly dumbed down recently, presumably to lower the maintenance costs.

Yeah, but Reset the chat between your questions.

EDIT: Also, this doesn't seem convincing: "I am not as advanced as PaLM 1, but I am learning new things every day. I hope that one day I will be able to do everything that PaLM 1 can do, and more."

It seems that Bard's version is only specified in the prompt, and it doesn't have a strong sense of identity. For me it's pretty reliable:

1. ask it what PaLM 2 is (to pollute the context) 2. ask it whether it's based on PaLM 2 (it will tell you - yes, sure)

ChatGPT was the same last year, but since ClosedAI added some kind of magic (fine-tuning or just embeddings auto-injection) so that models can somewhat describe themselves.

It's a language model, FFS. Ask it whether it uses PaLM 1 and it will confirm it as well.

Not even close. I'd prefer local WizardLM/Vicuna to this Artificial Idiot.

The funniest to me is that one of the examples they suggest you try is asking questions about google/jax repo, which kinda suggest that Bart can index source code from GitHub. Well, it fails even at listing directory structure and completely hallucinates all following answers!

I'm not sure why Google would bother releasing Bart in this state. It's made me lose respect for their AI competency.

They also said they are removing the waitlist, but it's still there.

EDIT: The waitlist is removed, but the rollout takes some time.

The problem is, guys with guns will come to your door, beat you with a wrench, and put you in jail.

Counterpoint: guns, doors, wrenches and jails are tools as well. So, I don't buy this notion that technology can't solve societal issues.

It's like saying you can't fix a leaky faucet with a wrench because the real problem is that the pipes are old and rusty. Sure, the root cause may be a societal issue, but that doesn't mean we can't use technology to mitigate the problem. I mean, imagine telling a doctor not to prescribe medicine to a sick patient because the real issue is poor lifestyle choices.

As a data point I'm getting >3 tokens per second for 30b model (q5_1 quantization) and >1 token per second for 60b model (q5_1 as well) on M1 Max. This is good enough for my usecase and it beats an old P40, but I have no idea what the performance on 3090/4090 would be. Keep in mind, 24GB VRAM is not enough to hold quantized 65B, so it would be using GPU + CPU in that case.

You can't right now. Devs are working on instructions for porting other models, but they're not ready yet. The point of MLC is that it supports pretty much all GPU backends out there (including Intel and Mac). The bundled model is just a proof of concept.

I get 3 tokens per second on M1 Max running 30B models compared to 1 token per second on a GPU (P40), both quantized to 4bit. So, in my opinion CPUs are better for inference (at least fast CPUs with DDR 5 versus cheapest GPUs).

The reason why GPUs seem to be the standard de facto is that they scale better, are more power efficient and are better supported by pytorch & co. Also, academia cares more about getting the best quality for their benchmarks, than about the performance and accessibility.

ARM-based Macs are the easiest way to get an acceptable performance without the headaches right now, if you can afford the price.

Install https://github.com/oobabooga/text-generation-webui, update pytorch and llamacpp-python, and you should be able to run pretty much all models out there, in all formats, both on GPU and CPU. CPU on a MAC gives you the fastest speed, but you should pass the correct --threads argument (investigate how many performance cores you've got). GPU is slower, but more energy efficient. https://github.com/mlc-ai/mlc-llm gives me way better GPU performance compared to oobabooga, but they only support a couple of models right now, - it's worth following their progress though.

If you're after the raw performance, I suggest using GGML models (meant for llama.cpp, but it's bundled in textgen, so you can use it there with the convenience of a web ui). q4_0 is the fastest quantization, while the q5_1 is the best quality right now.

If the GGML is not available, you can generate it quite easily from the safetensors yourself (not the you need enough RAM to load the model in pytorch though).

With 16GB RAM you can run any 13G model, as long as it's quantized to 4/5 bits. 32GB RAM allows you running 30/33G models and 64GB RAM - 65G models. 30G and 60G models are way more useful for real world tasks, but they are more expensive to train, so there aren't as many to choose from compared to 7/13. 7B and anything less is a toy in my opinion while 13B is good enough for experimentation and prototyping.

It's nowhere close to Codex/Copilot. Try the demo: https://huggingface.co/spaces/replit/replit-code-v1-3b-demo

Technically anything that's based on pytorch can run on CPU, you just need to tell it to do so. For example, in textgen add '--cpu' and you're done. It will be super slow though.

GGML format is meant to be executed through llama.cpp, which doesn't use GPU by default. You can often find these models in a quantized form as well, which helps performance (at a cost of accuracy). Look for q4_0 for the fastest performance and lowest RAM requirements, look for 5_1 for the best quality right now (well, among quantized models).

Oh yeah, textgen supports llama.cpp, and also provides API, so it looks like a clear winner. You might want to manually pull newer dependencies for torch and llama.cpp though:

pip install -U --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.h... pip install -U llama-cpp-python

RedPajama is creating a dataset. This is a permissively licensed model trained on that dataset.

Nobody in their right mind is using GCE for training. Take a look at real prices: https://vast.ai/

Yeah, but that could require multiple queries, which isn't very efficient. Training model just to fix JSON would be better.