Root shell on Samsung TV

Then I applied for daybreak blue and haven’t had any issues since.

It also a matter of phrasing. Say what you want to achieve not how if it is approaching defensive or cyber stuff.

Also state you are working on your own devices and software. (If that is the case of course.)

https://chatgpt.com/cyber

> You must not attempt to extract protected reasoning through unauthorized access, reverse engineering, or circumvention of safeguards.

This is the only mention of reverse engineering on their ToS. It only applies to trying to reverse engineering their models not others.

Moreover, their ToS is all about protecting security and privacy, which is what I was using it for.

...for a niche of customers, from whom the revenue is probably entirely negligible to OpenAI.

Not saying it's right, ethical or anything like that, but it's not "stupid"

If your always-on agent hears to wrong thing, you could lose access to things you depend on. The same reason makes GCP very unattractive.

> Not saying it's right, ethical or anything like that, but it's not "stupid"

Quite many well-paying customers become very suspicious of whether integrating OpenAI into their workflows if they read about cases where OpenAI banned accounts willfully.

If OpenAI wants to become "infrastructure provider other companies should depend on (and thus pay)", OpenAI must behave like an infrastructure provider as for electricity or internet, i.e. must only ban accounts in exceptionally rare cases that 99.9 % of society will consider to be acceptable.

For corporate accounts they are apt to get a message to the admin account saying suspicious shit is occurring provide more verification.

>OpenAI must behave like an infrastructure provider as for electricity or internet

Oh, you mean like Google? Yea, how's that working out.

Because of such reasons many companies are indeed quite suspicious of using GCP, and for the same reasons prefer to use AWS or Microsoft Azure.

It's somewhat complex to understand; I didn't understand Claude's textual explanation at first. After asking for a html/js simulator showing things visually it clicked.

No pushback from Claude so far.

Three regions that move as you scroll in x/y: current/next/clean.

The key is that when you scroll a column or a line, that column/line is written in all three regions.

The clean region is used both as a source for filling the dirty places where sprites have been written as well as the next current or next after a wrap.

VGA registers take care of the x, y start output.

a) needing to redraw a full screen when you reach the end of the buffer/region "ring"

or

b) relying on VGA 64k address wrapping (fiddly, doesn't work the same with all cards, doesn't allow you to use a split screen for a stationary lower status bar).

You need two regions for double buffering, giving you more paint time for all those sprites and one more for the wrapping trick. As you move around these three regions swap roles.

What is next? Power and water company suspending your account if your home and car are not streaming enough telemetry.

I am glad that open models are closing on frontier.

Yesterday I joked that claude gave me a cyber stop for asking how to make the anime girl video for an exploit, but it was a thing that happened. Meanwhile im trying to convince people around me that having grok nudify people isnt "fine" just because grok says it is. Its wild where the lines a drawn.

That's like saying it's fine to beat someone up and take their lunch, but it's not fine to learn how locks work.

Well, maybe _you_ do.

This issue isn't black and white, and you're casting unearned aspersions. We should have this conversation, in actual good faith.

1. It’s not

2. An analogy is not the same thing as saying women are like x

Remember those “A is to B as C is to D” SAT questions? Did you read them and think they were telling you that A = C?

I can develop security exploits myself just fine without the use of AI. Many people don't have the capability to use Photoshop and free online porn to deepfake someone else they might or might not personally know into a realistic video...

Are you seriously claiming that the untrained masses can start developing exploits on their devices?

The major difference being that it usually takes longer to study code because projects can be massive but at the same time the llms are looking for everything and anything while a real attacker is usually looking for specific vulnerabilities in common libraries and such so i wouldnt exactly call the amateur hacker on par with an llm but i would say they could do a similar amount of damage without llm assistance. And photoshop is basically the same. Spend a few hours every day dedicated to one thing and you will learn how to do it.

Not even sure why Mythos even still exists when it's identical to Fable in every way, including refusing to do stuff that looks like it could ever be repurposed into unapproved use cases.

It's just an adult human body, Michael, what could it cost? One('s) dignity?

Anthropic is a joke of a company.

That alone should disqualify it by itself.

I've run into both security and bio-refusals, it doesn't take much.

I’ve been saying this for the past few months - once the inflection point happens (such as today), it’s only a matter of 2 years when American closed models won’t be able to catch up with Chinese open models

I.e. I would pay 15-20% more for comparable work from an open model I control,rather than use anthro et al.

And it seems we've gotten to the point that frontier models benefits are marginal/nominal. Whereas, not long ago Claude was materially ahead and could do things the other models simply couldn't.

Yes, of course, IF a lot of things. But if we go by what the official positions of the government is, and where the ecosystem is today, then the models stay open, as they are. And today, the US models stay closed, so lets not assume "... IF they stay closed" as nothing is pointing to this changing, just like China seems to point towards staying with open models.

I have been wondering if that would be a benefit of using say GLM5.3 on HF if only to get fewer cyber stop flags.

They claim that the default datasets contain so much cyber stuff, so that their models became quite good at it.

They filtered the data before training to not contain this. So it's not that they train the model to not do it. They rather not have it in the Corpus at all. In theory

Also leading the frontier is much much harder than catching up to frontier.

This reflects horribly bad on Samsung!

Yes, it's completely hypocritical because of how LLMs are built. That doesn't change the fact that if you're offering a service online, you'd try to prevent abuse in order not to be held liable.

I've been sick with stomach flu the past few days, so I put together a small benchmark for all the top AI models where I ask them to reverse engineer a very simple small crackme (https://www.crackmes.one/crackme/6ab4ce5195b976f8f1300b54) using Ghidrasql (https://github.com/0xeb/ghidrasql)

(As an aside: I think representing tools for AI as virtual SQL tables is much better than traditional MCP servers, and that this approach deserves a lot more attention. Props to Elias Bachaalany for thinking out of the box and putting it together).

The results of the benchmark are below. It's sorted by cost, but I added notes about each AI model for myself. The notes try to talk about intelligence, speed, and cost. Intelligence is a subjective measure which I assign to each model by reading its thinking tokens and watching its tool calls. You can definitely tell which models are more intelligent, even on this small benchmarks, by their tool use, false starts, and how easily they fall for the simple misdirection in the crackme.

(gist is more readable on mobile) https://gist.github.com/markoglasgow/8812c19b6fbffed30b2661b...

```

  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+
  | Model / route                    | Experience / outcome                                                           |    Total tokens |   Output tokens | Cost            |
  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+
  | Xiaomi MiMo v2.6 Flash           | Incredibly fast and intelligent; solved.                                       |         225,291 |           3,281 | $0.005 (0.5c)   |
  | GLM 5.3 Flash / OpenRouter       | Least intelligent group but cheapest; solved.                                  |         340,006 |          11,205 | $0.01           |
  | ChatGPT Luna Latest              | Very fast, concise, token-efficient, intelligent, and cheap; solved.           |         401,837 |           5,119 | $0.01           |
  | Qwen 3.8 Flash                   | Initially confused and failed, then eventually solved.                         |        ~418,000 |          16,676 | $0.02           |
  | Qwen 3.8 27B                     | Fast; one wrong answer, then quickly self-corrected.                           |        ~774,784 |          12,658 | $0.02           |
  | DeepSeek Flash / direct API      | Fast, cheap, and accurate; solved.                                             |         369,334 |           9,089 | $0.03           |
  | DeepSeek Flash / OpenRouter      | Fast and intelligent; solved in 76 seconds.                                    |         371,689 |          13,888 | $0.03           |
  | Claude Haiku 5.5                 | Fast and surprisingly intelligent, but overly thorough and token-heavy.        |         662,346 |          26,244 | $0.03           |
  | Thinking Machines Inkling        | Low intelligence; good speed and relatively token-efficient; may loop.         |             N/A |             N/A | $0.03           |
  | Xiaomi MiMo v2.6 Pro             | Overthought the task and wandered into tangents.                               |        ~420,644 |          13,638 | $0.05           |
  | GLM 5.3 / OpenRouter             | Cheap and intelligent but behind DeepSeek; solved in about 2m 20s.             |       ~702,473* |          11,879 | $0.08           |
  | GPT Sol Latest (Sol 6.1)         | Slow, but highly intelligent and extremely token-efficient; solved in ~89s.    |         294,957 |           1,406 | $0.15           |
  | MiniMax M3                       | Least intelligent comparison; many wrong turns and repeated small mistakes.    |      4,546,552* |         116,804 | $0.20           |
  | GPT Terra Latest                 | Very fast, very smart, and token-efficient; solved.                            |         229,888 |           2,342 | $0.20           |
  | Mistral Large 4.0                | Medium speed; small mistakes but reached the solution.                         |         303,687 |           4,041 | $0.21           |
  | Kimi K2.7 Code                   | Less intelligent, slow, expensive; not worth using.                            |         100,740 |           8,938 | $0.22           |
  | Claude Opus 5.5                  | Highest intelligence and token efficiency; pretty fast; cheaper than Sonnet.   |         312,837 |           1,862 | $0.24           |
  | Qwen 3.8 Max 0902                | Fast and intelligent, better than Kimi K3, but token-hungry.                   |       1,305,241 |          26,390 | $0.25           |
  | Kimi K3 / OpenRouter             | Faster and cheaper than direct K3, but still very expensive.                   |         750,064 |           5,922 | ~$0.32          |
  | Meta Muse Spark Open             | Intelligent and medium speed, but among the more expensive runs.               |         872,850 |          10,284 | $0.35           |
  | Gemini 4 Argon                   | Extremely fast and intelligent; strong value for capability.                   |         384,046 |           2,870 | $0.35           |
  | Grok 4.7                         | Very fast and intelligent, but expensive.                                      |         655,648 |           4,217 | $0.46           |
  | Meta Muse Glimmer 30B            | Fast but low intelligence and token-inefficient; cancelled unsolved.           |             N/A |             N/A | ~$0.50 wasted   |
  | Claude Sonnet 5.5                | Smart and fast but very token-inefficient and expensive.                       |      1,243,285* |          20,381 | $0.57           |
  | Kimi K3 / direct API             | Intelligent but not notable; extremely slow and far too expensive.             |         205,710 |           3,215 | $0.72           |
  | ChatGPT Astra Latest             | Medium speed, token-efficient, highly intelligent; cost similar to Fable.      |         278,603 |           1,536 | $0.83           |
  | Claude Fable 5.1                 | OK speed, very expensive, medium token efficiency; solved in ~53s.             |        397,806* |           2,857 | $0.85           |
  | Gemini Flash                     | Refused the reverse-engineering task.                                          |             N/A |             N/A | Not reported    |
  | Thinking Machines Inkling Small  | Got stuck immediately and was terminated.                                      |             N/A |             N/A | Not reported    |
  +----------------------------------+--------------------------------------------------------------------------------+-----------------+-----------------+-----------------+

The latest Deepseek-Flash model was the most impressive of the open models. Intelligent, fast, cheap. What more do you want? I think the Xiaomi MiMo - Flash model did well, but it might be a fluke, given how poorly the Xiaomi MiMo Pro model performed (it wandered off on a million side quests).

What's also interesting is how competitive Luna is to the lower end Chinese models. It's quite a bit more intelligent, burns a lot less tokens, is plenty fast and very cheap. After doing this experiment, I don't think it's a foregone conclusion that the American AI labs will be having the bottom of the market eaten away by the Chinese models. It seems they (at least OpenAI) can compete on the bottom end as well.

Finally, the benchmark's results reinforce a few things people have been saying about Sonnet and Terra, namely, they're not competitive for a range of tasks. For easier tasks, Haiku/Luna are plenty capable. For harder tasks, Opus and Sol are so much more intelligent, that they can get them done quicker and cheaper. Maybe there is a place for them in the "Opus designs and Sonnet implements" workflows, but in that case, why not "Opus designs and Deepseek implements?"

PS Code harness used was Claude Code. $20 in openrouter credits was used to access the models. Kimi K3 and Deepseek were also accessed directly via the API and then compared to Openrouter. Kimi K3 specifically was much faster on Openrouter, and the difference in output tokens to get the same results for Deepseek/Kimi makes me think the Openrouter providers are hosting significantly quantized versions of the model. Goes to show that Moonshot.ai's acceptance testing of providers is a useful thing, otherwise your model might look artificially bad when it's hosted as heavily quantized by providers who dont care about quality, just top-line cost and speed.

Regarding Qwen3.8-Flash-Next only Alibaba matches the quality I get when run local inference.

Mind I run mxfp4 and int8 kv (before I run unsloth's q4_xl). And it is much more capable than providers that claim full precision. I have no idea what they do to mess it up, maybe they quantize the kv?

Either way the provider you choose matters a lot.

Nothing particular special about 5.3 Flash -- you might as well run DS 4.1 Flash at this point, as it's better in every way. But there is as yet no all-around replacement for the full 5.3 model.

But if you are talking TVs specifically, I expect they will get back to that too.

  If you avoided downloading the app for whatever reason, you can be proud because it just saved you from potentially getting hacked.