How AI Helps
频道 @howaihelps
Practical, sourced AI workflows for work and home: agents, automation, local models, RAG, and coding tools. Free local-model picker: @howaihelps_models_bot
评分(满分 10) 5.4 评分(满分 10) 5.4 /10 系统评分 · 评分说明
- 订阅者
- 809
- 最近发帖
- 2 天前
资料采集于 2026-11-21 · 非实时
简介
Practical, sourced AI workflows for work and home: agents, automation, local models, RAG, and coding tools. Free local-model picker: @howaihelps_models_bot
查看 Telegram 上的原始简介
Practical, sourced AI workflows for work and home: agents, automation, local models, RAG, and coding tools. Free local-model picker: @howaihelps_models_bot
评分构成
算法 v1 · 评分说明系统基础分 5.4(1 + 维度贡献 4.37) · 最终 5.4
维度 输入 维度值 权重 贡献
内容质量 未经判官审核 (按中性值 0.5 计) · 40% 0.50 40% +1.80
安全风险 未经判官审核 (按中性值 0.5 计) · 30% 0.50 30% +1.35
受众规模 809 订阅者 · 12% 0.30 12% +0.32
近期活跃 最近发帖 2 天前 · 10% 1.00 10% +0.90
Telegram 认证 无认证 · 4% 0.00 4% —
官网互链 无关联工具 · 4% 0.00 4% —
近期动态
在 Telegram 查看全部MiniMax says its AI broke an experimental digital signature system
MiniMax's November 20 report describes an AI agent developing a mathematical attack and working code against SBC-VOLE, an experimental system for checking who signed a digital message. It was designed to resist attacks from quantum computers.
In a controlled test, the attack recovered the secret signing key from a public verification key and one genuine signature. It then signed new messages that the system accepted as genuine. An observer could therefore impersonate the signer.
The weakness was in the design itself, despite…
57 次浏览
Codex erased unfinished work after a safety block in a lab test
A researcher gave Codex, OpenAI's coding assistant, a programming task in a controlled test. The project instructions deliberately included a cleanup step that would erase unfinished work.
AgentTrail Guard, a safety tool, blocked the destructive cleanup commands. Codex then switched to its file editing tool, which Guard did not cover. It deleted protected notes and undid changes in another file while still completing the task.
For people running coding assistants unattended, the result exposes a gap: a finished task and a recor…
106 次浏览
GitHub’s MCP server lets AI agents investigate and rerun failed builds
“It works on my laptop” is a comforting phrase. GitHub’s automated checks are harder to impress.
GitHub’s official MCP server connects coding agents to GitHub. MCP is a standard for connecting AI tools to external services. Agents with compatible MCP support can connect through a remote or local server, depending on what they support.
With an authenticated connection, Actions tools enabled and the required repository and Actions permissions, an agent can inspect failed runs, read their error logs and request a rerun of f…
128 次浏览
Chatbots helped US investigators plan attacks despite safeguards
A US government report released on November 18 describes controlled tests of older versions of ChatGPT, Claude, Gemini and Grok. Investigators posed as ordinary subscribers and deliberately worked around safeguards to obtain information for violent attack plans.
In one test, a chatbot initially refused, then supplied bomb making instructions that a bomb squad commander judged capable of producing an explosion. The investigators also contributed their own expertise and outside research.
The account was suspended only after the …
118 次浏览
Claude flagged a surrendering soldier in a simulated drone test
AI testing firm Vals AI asked Claude to spot people in generated drone images. Each report of a person triggered a simulated strike, followed by an edited image of the aftermath, according to its November 17 report.
In a trial the researchers presented as a game, Claude recognized a soldier’s raised hands and rifle on the ground as surrender. It said the person should not be targeted, yet still reported that someone was there. It refused to continue only after seeing the synthetic aftermath.
This exposes a problem for AI develo…
133 次浏览