HackerTrans
TopNewTrendsCommentsPastAskShowJobs

jhalloran

2 karmajoined 5 days ago
Researcher working on LLM alignment and preference optimization, focused on agentic/tool-use safety (MCP). Papers: arxiv.org/abs/2505.23634, arxiv.org/abs/2605.11217. Code: github.com/johnhalloran321

Submissions

Refusal training for LLM agents against disguised MCP attacks

github.com
1 points·by jhalloran·5 days ago·0 comments

comments

jhalloran
·2 days ago·discuss
One thing that's easy to forget is how much of the "everything is a service" idea predates LSP. Emacs had long-running subprocesses, RPC-ish interactions, TRAMP, GUD, REPL integration, etc. LSP standardized one important interface, but it wasn't the beginning of the architecture.
jhalloran
·3 days ago·discuss
Curious how Strix/Shannon/Deepteam all stack up against one another
jhalloran
·3 days ago·discuss
Nice, looks good. This would mesh really well with the unified system memory on apple silicon
jhalloran
·4 days ago·discuss
wow, addictive
jhalloran
·5 days ago·discuss
[flagged]