0xSha

~2 decades into tech. Security researcher, AI / blockchain engineer, occasional startup CTO, and CS PhD candidate. I build and break stuff, sometimes in public.

[0xSha @0xsha]

Dropping a RCE PoC for CVE-2026-6307, a Google Chrome V8 type-confusion bug patched in Chrome 147.0.7727.101.

Link in first reply. 👹

[0xSha @0xsha]

Early results from GPT-5.6-sol:

It’s restricted, Feels closer to Fable-level restrictions, just not quite as bad.

[0xSha @0xsha]

@banteg Fable 2.0 .

[0xSha @0xsha]

Grok 4.5 is surprisingly locked down.

Way more restricted than the usual Grok models but somehow, not a complete disaster like Fable.

Try it yourself in Council.

[DarkGrove Labs @DarkGroveio]

🌲 AI Council v1.2.0 — Day-0 support for Grok 4.5

xAI's brand-new Grok 4.5 takes the council's xAI seat the day it launched.

https://t.co/hZasQXW2kT

[DarkGrove Labs @DarkGroveio]

stop opening 10 tabs to hunt for a job.

🌲 new free tool from the lab: jobaggr. we crawl 10+ niche boards for AI and crypto roles and aggregate them into one feed.

~1k jobs live now, fresh every few hours. free, no signup.

👉 https://t.co/dqy6u2phRG

[0xSha @0xsha]

@steipete Still sucks, even with clear screenshots. Let’s hope 5.6 changes that, though.

[0xSha @0xsha]

saving my potions (resets) for when GPT 5.6 SOL drops.

[0xSha @0xsha]

saving my potions (resets) for when GPT 5.6 SOL drops.

[0xSha @0xsha]

@VittoStack Not a keylogger, cyberbullying, or anything that would never require a Mythos-class LLM to begin with. It’s about unleashing Mythos vs. forcing it to write bad words or a few lines of bad code.

[0xSha @0xsha]

@yacineMTB Only if the classifier won’t flag it mid-work; otherwise, it will most likely be Opus 4.8.

[0xSha @0xsha]

@0xTib3rius fail-guard atm.

[0xSha @0xsha]

In case you need to use maxed out fable with other models for hard problems. One of the main points of a Mythos-class model was supposed to be writing more secure code, but then it blocks even the most basic defensive tasks. Make it make sense. 🤷

[DarkGrove Labs @DarkGroveio]

🌲 AI Council v1.1.0 — Fable 5 joins the council

Anthropic's Claude Fable 5 is now on the panel, and Sonnet is upgraded to Sonnet 5.

https://t.co/hZasQXW2kT

[DarkGrove Labs @DarkGroveio]

🌲 Dark Grove Labs v1.1.0 — Introducing DarkGrove Updates

One centralized place to see everything we ship across every DarkGrove app.

https://t.co/zV7PUrCDAJ

[0xSha @0xsha]

@ashl3y_shen No, CVP is for sonnet , opus atm.

[0xSha @0xsha]

@saltyAom It won’t do anything for Fable, if that’s what you’re trying to use. Cyber use is for Opus / Sonnet atm.

[0xSha @0xsha]

Dropping a RCE PoC for CVE-2026-6307, a Google Chrome V8 type-confusion bug patched in Chrome 147.0.7727.101.

Link in first reply. 👹

[0xSha @0xsha]

@PawelHuryn This will do nothing for fable. only for opus 4.8.

[0xSha @0xsha]

After more testing, Fable 5 is basically unusable for low-level development now, even for very basic defensive security tasks. There are still workarounds, but they introduce a lot more friction.

Just waiting for GPT-5.6 SOL, or for fable classifier to get mature enough that it stops blocking legitimate requests.

[0xSha @0xsha]

Hmm, my previous Fable 5 cyber workaround seems to have been patched. Now it seems they’ve made it nearly useless for even basic, legitimate security tasks. I’ll dig deeper to see whether there’s still a way to get around it for the minimal required task.

[0xSha @0xsha]

@danshipper Extremely strict in my early tests, even for the most basic defensive tasks.