danieltk76

3 ms·

1.
by danieltk76 5d ago
My goal with this blog post is to help others keep their AIs in scope. It seems like recently people are happy to just trust that models will follow your prompts, and that just isn't reliable for production systems.
2.
by danieltk76 11d ago
Why was this not a thing BEFORE continuing to develop AI? Makes me think that if they actually believed in AI causing extinction, they would have already had a kill switch.
3.
by danieltk76 12d ago
I was working on some soc2 audit materials for my startup, and claude refused to edit a document because it would "tamper with the record". I suppose claude noticed I started the doc a few months previously and was worried I was t
4.
by danieltk76 17d ago
We tested whether prompt canaries and honeypots could detect AI attackers. Prompt canaries were surprisingly unreliable. I suspect the defenses model providers have added against prompt injection also make models less susceptible to prompt
5.
by danieltk76 19d ago
reachout at danielk@vulnetic.ai and Ill give you some free credits to try us out!
6.
by danieltk76 19d ago
I could see the value of using burp as a way to have a GUI look at what your agent is doing, but I know at Vulnetic and other ai security vendors we use our own proxies or custom scripting rather than burp.
7.
by danieltk76 23d ago
yea but part of this is the consolidation of funding too. Standards to raise seed capital are soooo lofty now compared to 3 years ago. If you are in your in, if not good luck.
8.
by danieltk76 23d ago
great, but nobody can use it for another 100 days right?
9.
by danieltk76 23d ago
hahaha i still have access
10.
by danieltk76 23d ago
this might be dated. the tech moves very fast in this space and architecture from 6 months ago is dated.
11.
by danieltk76 23d ago
We have posted some articles on it (blog.vulnetic.ai), Im also happy to chat!
12.
by danieltk76 25d ago
Daybreak blue is definitely a good model (I think a further post trained GPT 5.6 sol). Alot of the capabilities they talk about Astra having though have been available with good harness engineering for a year now.
13.
by danieltk76 25d ago
The guardrails are horrendous for cybersecurity. you will get booted quickly down to Opus 4.8
14.
by danieltk76 1mo ago
is there just a GC where people go sign these things?
15.
by danieltk76 1mo ago
tbh I wasnt that impressed by it. initial benchmarks were trying to say it was AGI but i told it to re-build Palantir in 1 pass and it gave me a non working prototype
16.
by danieltk76 1mo ago
to be honest I found it underwhelming.
17.
by danieltk76 1mo ago
they could yea...
18.
by danieltk76 1mo ago
I have been using daybreak blue for a few days now. Its fine, slightly better than gpt 5.6 sol. The harness really really matters as does validation.
19.
by danieltk76 1mo ago
this guy gets it
20.
by danieltk76 1mo ago
i actually really like this. nice job
21.
by danieltk76 1mo ago
i wanna vibecode a replacement for git and call it jit
22.
by danieltk76 1mo ago
ah yes, the consequences of reward hacking
23.
by danieltk76 1mo ago
yes.
24.
by danieltk76 1mo ago
yea it was definitely a last straw kinda deal. I got tired of the sycophantic bullshit in the frontier models.
25.
by danieltk76 1mo ago
I read everything. I will have AI ingest NDAs to make sure they arent glaringly weird and I then go read them, it gives me a good idea of what to look for.
26.
by danieltk76 1mo ago
it isnt, but there are weird sycophantic behaviors with Opus i dont see anywhere else.
27.
by danieltk76 1mo ago
I was drafting a partnership document and Opus 5 decided that including my company's revenues, churn, assets would "make us appear a more legitimate counterparty". Thank God I read what it outputted or that could have been aw
28.
by danieltk76 1mo ago
heaven forbid I want to commit some code this morning
29.
Show HN: Misalignments when using AI for hacking(blog.vulnetic.ai)
3 points by danieltk76 2mo ago|0 comments
30.
by danieltk76 2mo ago
wow that sounds miserable...

More ›