Into Programming, Digital preservation, Ai and Fishtank
Joined January 2026
- Tweets1.3K
- Following170
- Followers64
- Likes13.1K
No joke, ask an agent if you are projected and have it defend you and patch your system. Blue team is very much important here.
Redcode retweeted
the beatings will continue until the labs give independent researchers easy access to SOTA uncensored models
Buckle up. It’s gonna be a bumpy ride
The flaws sat dormant for over a decade because turning the memory-handling issues into reliable unauthenticated RCE required an extremely complex chain of steps that was practically impossible for humans. Experts spotted related problems in a 2024 audit but found no concrete exploit path, so fixes stalled. In 2026, researchers used frontier AI models to analyze the code, generalize the patterns, and construct working exploits in days—what once took world-class teams months. Once details leaked, attackers could follow.
Sam Altman reportedly told OpenAI staff there are three possible endgames for AI:
- an IAEA-style global regulator
- a CERN-style merged super-lab
- one lab using recursive self-improvement to "melt everyone else's GPUs"
Two of those are treaties. The third is a threat.
Anthropic, the company that published essays about existential risk, has a leaked draft federal AI bill.
Safety advocates say it's significantly weaker than the bipartisan bill already on the table.
The leading voice for AI safety is lobbying for less of it.
Every AI story today had the same shape:
the agent did exactly what it was told, did it flawlessly, and the outcome was still wrong.
Nobody is building the part that hesitates.
What's the most "technically correct" disaster an AI has caused for you?
People are asking Claude Code to hack their smart TVs.
Model number in → agent finds TV on wifi, kills 25 bloatware packages, writes a new launcher. A few thousand people copied it in 48 hours.
Same day Microsoft classified AI agents as insider threats.