HackerLangs
TopNewTrendsCommentsPastAskShowJobs

syumei

78 karmajoined 4 tahun yang lalu
Ph.D. student @columbia working on Security, Crypto, and AI/ML

Submissions

A Cursor Sandbox Escape Shows Why AI Agents Need Kernel Boundaries

medium.com
1 points·by syumei·3 hari yang lalu·0 comments

A Conflict-Free Multi-Agent Ensemble for Claude and Codex

medium.com
1 points·by syumei·7 hari yang lalu·2 comments

CurXecute: When a Slack message rewrites your editor's config and runs code

medium.com
3 points·by syumei·7 hari yang lalu·0 comments

Claude Code SOCKS5 Bypass Shows Why Egress Filters Belong at the Boundary

medium.com
1 points·by syumei·8 hari yang lalu·0 comments

Reached 440+ stars: Auditable sandbox to record what AI agents did

old.reddit.com
2 points·by syumei·9 hari yang lalu·0 comments

Auditable Workspaces for AI Coding Agents

medium.com
1 points·by syumei·10 hari yang lalu·0 comments

I built an anime-style UI for watching AI coding agents review each other's code

old.reddit.com
3 points·by syumei·12 hari yang lalu·0 comments

How Claude Code and Codex Sandbox Untrusted Code

medium.com
2 points·by syumei·13 hari yang lalu·1 comments

Two Heads Are Better Than One: Run Many AI Agents, Merge One Auditable Result

medium.com
2 points·by syumei·18 hari yang lalu·0 comments

How Claude Code Broke My Git Worktree

medium.com
3 points·by syumei·24 hari yang lalu·0 comments

A Git/LFS-backed unified tool-output format for AI agents, with 95% token cut

old.reddit.com
2 points·by syumei·bulan lalu·0 comments

Claude Code and Codex can have real-time conversation via Git

medium.com
116 points·by syumei·bulan lalu·79 comments

[untitled]

1 points·by syumei·bulan lalu·0 comments

Git-based cache saves 50% on token usage

old.reddit.com
34 points·by syumei·2 bulan yang lalu·1 comments

[untitled]

1 points·by syumei·3 bulan yang lalu·0 comments

[untitled]

1 points·by syumei·3 bulan yang lalu·0 comments

Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval

medium.com
2 points·by syumei·3 bulan yang lalu·0 comments

[untitled]

1 points·by syumei·3 bulan yang lalu·0 comments

[untitled]

1 points·by syumei·3 bulan yang lalu·0 comments

[untitled]

1 points·by syumei·3 bulan yang lalu·0 comments

comments

syumei
·7 hari yang lalu·discuss
Thank you so much!
syumei
·13 hari yang lalu·discuss
Claude Code and Codex both run untrusted commands. This post compares how their sandboxes protect your repo, secrets, and host machine.
syumei
·24 hari yang lalu·discuss
quite interesting post
syumei
·bulan lalu·discuss
Fair criticism. I don’t think the value is just "agents can send text to each other”; that part has many implementaton and design choices. The value of h5i I’m exploring is making the intermediate state reviewable: review requests, risks, handoffs, unresolved claims, associated prompts and AI-to-Ai conveersation, and final decisions tied to the branch/PR.
syumei
·bulan lalu·discuss
[dead]
syumei
·bulan lalu·discuss
[dead]
syumei
·bulan lalu·discuss
[flagged]
syumei
·bulan lalu·discuss
[dead]
syumei
·bulan lalu·discuss
[dead]
syumei
·3 bulan yang lalu·discuss
[dead]
syumei
·3 bulan yang lalu·discuss
I've been tracking real-world AI agent failures and incidents recently.

Things like:

prompt injection → goal hijacking

agents going rogue due to misalignment

unintended/unsafe tool use

It feels like we're starting to see repeatable patterns, not just isolated bugs.

I'm collecting cases + papers here:

https://github.com/h5i-dev/awesome-ai-agent-incidents

If you've seen interesting incidents, weird failures, or relevant research, I would love to add them.