varenc

The poster: https://cdn.saffire.com/images.ashx?t=ig&rid=OhioStateFair&i...

Note the pigs are numbered 1, 2, 1. And the gondola with a floating cable and another that terminates at a tent top.

It's interesting how much of the value here is providing the english Skill definitions that tell the LLM what to do: https://github.com/openai/codex-security/tree/main/sdk/types...

Some of approaches there could be useful in other contexts. OAI has the compute to experiment with different prompts and I'd expect these to be somewhat optimized.

You're ignoring the asymmetry with security. The attacker just needs one exploit chain, whereas the defender needs to block every avenue. Open access to models with no guardrails greatly benefits the attackers more than the defenders.

Imagine what a god-level hacking AI could do. It could find a full 0-click to root exploit chain in iOS. Attacker unleashes a worm that infects a phone, instructs that phone to send the same attack to all of its contacts, and then physically destroy the phone by turning off all thermal throttling. Might even be possible to make it catch fire.

Or find a remote exploit in Tesla cars and make their autopilot go on murdering rampages. (that one is from a movie)

the 'that happened' makes it too long for the HN submission title length limit.

Maybe simonw can suggest an alternative title that fits within the limit, that doesn't misrepresent the post.

Agreed. Big love for exe.dev. For quick low stakes projects, using their Shelly assistant is the simplest way to give an LLM a fully persistent VM with web services I've found.

Digression: You can link directly to a page in a pdf with a url like this: https://arxiv.org/pdf/2607.13335#page=27

The way source maps are used by dark net sites, and the reason it exists, are probably unrelated! I think it was just an early form of using image tiling for big efficiency gains. Now with HTTP2/QUIC you can request many small individual images without much overhead, but back the day there were huge gains by sending one larger image instead of 50 smaller ones. (There's also the pure CSS solution used with image sprites that's between the old web and the current era, but I digress)

great point! Though what about the default windows APIs for 'calling out to shell'. Like using `subprocess.run(...)` in Python. My suspicion is that works like cmd.exe?

edit: answered myself. Seems like subprocess.run works like PowerShell, unless you pass in 'shell=True', then it passes it directly to CMD.exe. And I'm guessing that's what Cursor is doing.

fun fact: source maps are used by some Tor onion sites ("dark net") as part of a captcha process without using JavaScript. If you present the user with an image, and ask them to click on a particular part of it, the server can recieve exactly where they clicked and validate if that's correct without using JS at all. (JS is a big no-no on Tor hidden network sites)

Same. Recently Ubiquiti has been putting more and more into the local API, so this should be getting easier to do. The Home Assistant UniFi integration just recently has started moving from private undocumented API endpoints to the newer public API endpoints.

In other news, Meraki has an AI assistant feature now.

Free Free is celebrating this DEA rule because they say they're exempt due to their use of natural kratom leaf, not synthetic: https://botanictonics.com/pages/consumer-education

I am skeptical though since I presume their capsules are much stronger than natural kratom powder. The linked YouTube video seems to say that Feel Free intentionally includes higher concentrations of 7-OH. So a bit confused.

I hope the Feel Free does experience the crack down. Seems like someone just needs to test the Feel Free product's 7-OH concentration level, and if it exceeds 0.05%, as it likely does, the ban will apply.

I doubt it. This is just normal Windows behavior. If you call out to a shell and have it run a `git` command the system has to decide where the git binary is. A quirk of Windows is that first it searches the current directory for a `git` executable, and then it searches directories defined in PATH. If there's an executable git.exe in the current dir, that gets ran. I assume Cursor's behavior is the same across platforms but this is only a Windows vuln because Windows picks binaries in the current directory before PATH.

A Windows user purely using the CMD shell, not Cursor, that checked out a malicious repo, went to that repo, and then ran any `git` command in that repo could be hacked the same way.

On Windows, if your shell was in a malicious repo's directory and you ran any "git" command, wouldn't the behavior be the same? Binaries in the current directory are considered before PATH. This seems like a broader issue.

That said, Cursor should at least have some "trust this directory?" dialogue.

This is false. Sea water is >3% salt and human kidneys can't produce urine with that much salinity or greater. Since they need more water to extract that much salt, the net effect is dehydration. This rower probably had a desalination machine, or just a big reservoir of fresh water.

(Though there might be some obscure edge case, and if you're about to die of dehydration that a little bit of seawater will buy you a minuscule amount of more time? doubt it)

persistent.oaistatic.com for OAI software. Not just a subdomain, but a whole separate domain, making it easy !

Wow, stories like this make me happy that I block all ChatGPT/Claude/Codex/etc updates by default, and only selectively update on demand. I do this just by setting a network rule that blocks their update check. (fortunately updates are still served on a different domain than regular usage)

I would love it if a grocery store, or Uber, or Amazon, made it easy to see my yearly spend. Instead them to intentionally obscure this and try to prevent you from seeing a big number.

Kept trying to use the demo, but despite giving it microphone permissions and seeing that the microphone was in use, couldn't get the demo to 'hear' me at all.

You could also say life is too short to waste time retying inferior drawstrings. If learning the knot takes 5 minutes, it seems like it could easily save you time overall.

I don't think that quote from the article is disagreeing with you at all. Like you said, we don't have a cohesive definition or test of consciousness, so research like this doesn't say anything about if this is or isn't similar to human consciousness.

I would guess Anthropic included that sentence to make it very clear they're not claiming human-like consciousness, and dampen journalists writing headlines like "Anthropic discovers their AI thinks just like humans and may be conscious".

edit: later in the article they even more explicitly agree with you

> Our experiments don't show Claude can have experiences, or feel things in the way humans do—in fact, it’s unclear whether any scientific experiment could prove this to be true or false

This project is a fork and all it does is add iOS support. Commits by the repo fork owner begin 19 hours ago, so very plausible those are done with Fable. But I agree it's very unclear to me if adding iOS support was some task only Fable could accomplish over prior models.

Carplay isn't about entertainment for me. It's just a vastly superior and convenient way to view the same map you can see on your iPhone.

how so? Is it a class system that only Raytheon employees can work on cruise missiles, not the average citizen?

(edit: not that these models are equivalent to missiles oc)

That's a different and much more boring type of cheating. The interesting part of the METR report is that the model is hacking the evaluation environment, not that some AI model provider is hardcoding answers to known evaluation questions. (which wouldn't require the model to cheat/hack)

From my cursory reading of the pre-print, it didn't seem like the entire scroll has been read? Or rather, you read everything that can be read, but part of the original have been lost? (Like columns 1-4).

VERY COOL WORK!

It's false that the 'entire Herculaneum scroll has been read'. Much of the scroll has been lost. From the preprint, columns 1-4 lost, and then margins on other columns are also lost.

   Col. 5: "… the similar …"
   
   Col. 6: "… impulses …"
   
   Col. 9: "… so far as … this or to have … that …"
   
   Col. 10: "… that befits on the whole still … there will be fear and … the great and long …"
   
   Col. 11: "… and the impulse … For/towards each of these things in this way … we are by nature … and for/towards the fulfillment of these things that … seem …"
   
   Col. 12: "… to men and beasts … And above all, each of the most common things constitutes these … For, [necessity? necessary?] …"
   
   Col. 13: "… natural … therefore also … according to the … this … will be found, and lives will make no progress whatsoever, as we have no need for either pleasure or pain. In the same way, also …"
   
   Col. 14: "… and thus lacking … I want to say … common … accomplished … to lack … and … on the right parts towards the left ones. There is an excess in the impulse …"
   
   Col. 15: "… and of all similar things. For, according to this kind/category, according to which impulses exist by nature, there will be that which lacks nothing, so that one seeks nothing more, but completes in every respect as …"
   
   Col. 16: "… they approach completion. Moving from these things to … [λόγος?], it [τέχνη?] accomplishes within us all that pertains to it, even though it cannot fully complete nature. And it allowed …"
   
   Col. 17: "… we will inquire into something, but we will not grasp it, if in some way we depart from ourselves and from our own nature, and besides, in the same way as the remaining arts may be said to be perfected in one respect, but to be deficient in practical wisdom in another respect…"
   
   Col. 18: "… being that practical wisdom … and to be about it. This [sc. λόγος] concerning the mechanical arts seems to me to be very distant from such a [conception?], and to have the technical fulfilment that is, so to speak, lame and something of such type lacking, and concerning the …"
   
   Col. 19: "… need none. Having certainly strained ourselves to the utmost through research and learning, we will no longer be inferior to them in any respect, accomplishing in like manner the things that befit them and possessing the same practical wisdom as they …"
   
   Col. 20: "… to happen. And such being the goods for us, even from the opposite evils there will be neither anything good—let alone beautiful—nor anything bad—let alone ugly—nor happiness …"
   
   Col. 21: "… being greatly wise and celebrated and … to praise … as according to the eulogies …"
   
   Col. 22: "… still … Aristocreon … to possessed things …"

I think OP meant converting handwriting to text, not identifying a person based on their handwriting style! (but that sounds quite interesting)

As long as thinking blocks can't make tool calls, I don't really see the exfiltration risk.