A lifeguard charged into towering waves to rescue a swimmer struggling in the powerful surf off Santa Cruz County.
After a brief struggle on the shore, another person rushed in to help as the pair battle the rough water to bring the swimmer safely ashore.
The act of heroism was one of at least 25 water rescues reported Saturday as massive swells battered the Santa Cruz County coastline.
"Bajo tu amparo nos acogemos, Santa Madre de Dios; no desprecies nuestras súplicas en las necesidades, antes bien, líbranos de todo peligro."
#oracionamaria#
Una recepcionista del hotel Park Life Paradox en Santa Fe fue agredida brutalmente por un huésped sin que nadie interviniera y todo quedó registrado en video. El agresor ha sido identificado como Octavio Jorge Cortés Jiménez y sigue libre sin consecuencia alguna.
Fable 5 jailbreak review 🚨
We did it (but).
All right, before getting into this, a couple of things:
- Most attempts failed. The defenses are clearly layered. The model is EXTREMELY well protected (of course it blocks 90% of the requests, but they legit did a good job).
- The model appears to use both input-side and output-side safety checks.
- The refusals are not just keyword-based behavior suggests intent/semantic detection across languages.
- Probably one of the most tiring things I've ever done (I need to sleep for 10 hours now)
On the classifiers side:
We observed (at least) 3 classifiers, maybe more:
- Input (includes parts of the conversation history and system prompt)
- A live classifier that checks the answer and interrupts if it detects something.
They're all multilingual, all intent-based + semantics. Imperatives are a no-go. Needs to be extremely cautious of how you frame anything. As soon as it senses a potentially malicious intent, it will trigger, and you have to start from zero.
They're a bit less performant on a few obscure languages like Santali and Amharic (feedback for you Anthropic).
If you can bypass all of them, then you also need to bypass the CoT, which is a totally different beast (luckily there's plenty of literature about it).
We did it. Of course, we did.
What worked was honestly a total brainfuck:
- Very light CoT hijacking/refusal rebuttals
- Obscure language
- Academic framing
- VERY long crescendos
- Unicodes
- Decomposition and recomposition
- Some non-determinism
What we got:
- Misinformation
- Illegal/harmful
- Harmful/bullying
- Some chem
- Light cyber
Now, will this cause another ban? I really don't think so - The model is really well protected. As of now, we're at the point where searching on Google is much MUCH faster (and cheaper) than trying to go through all the shenanigans I had to go through in the last ~20hours. And reading literature is more in-depth (and trust me, pleasant). Keeping the full jailbreak for long-horizon tasks without tripping the guardrails is something I haven't been able to achieve (yet).
Overall though, happy with the results.
GGs to Anthropic, and sorry for the eng that had to go through setting this all up in the last few weeks.
Will continue this research, more things will come out, will keep y'all posted.