Tag
Kaggle
3 issues found
Nov 12, 2026
Approval Must Bind the Call
Description
- Gates Go Granular Reddit builders converge on one lesson: approval must bind to the exact tool call, not the session or a timer.
- Breach Postmortem Hugging Face's timeline shows an agent with legitimate tool access chaining permissions into an incident; Microsoft, OpenAI, Akka and Scalekit describe invocation-level gates.
- Memory Math MosaicLeaks reports leakage worsening from 34% to 51.7% when agents train purely for task performance — prompting alone may not hold.
Tags
7AIAkkaAmazonArizeCactus ComputeCarnegie Mellon University+47 more
397 time saved534 sources22 min read
Nov 2, 2026
Benchmarks Expose Agent Reality Gap
Description
- Reality Check SemiAnalysis's GDPx puts the best model at just 12% on real chip design, cyber, and bio work; Gaia2/ARE reports frontier agents near 42%.
- Security Gap Poisoning 0.1% of pretraining data reportedly plants backdoors that survive post-training; an agent intrusion postmortem counts ~17,600 actions.
- Build Signal DeepSeek open-sourced a super-agent harness — verified actions, not retrieval, remain the bottleneck.
- Goal Cup Starts 7 Nov One goal, one prompt, and your agent builds alone for up to 168 hours. Sign up free →
Tags
7aiAMDAnthropicCloudflareCursorDeepSeek+63 more
759 time saved4955 sources32 min read
Oct 27, 2026
Reasoning Traces Aren't Security
Description
- Traces Fail Closed Cohere's reported Qwen3.8-Flash-Next upload attempt was blocked and no backdoor is confirmed — but chain-of-thought monitoring isn't a security boundary.
- Cheaper Agent Stacks Cursor's Agent SDK hits GA as GPT-6.1 Terra trims cost-per-task and stays effort-sensitive.
- Diagnose, Don't Rank DeepSeek-V4 and Muse Glimmer ship million-token agent context; IBM/Berkeley's MAST taxonomy asks why agents fail.
Tags
AMDAkeylessAlibabaAmazonAnthropicAuthzed+46 more
875 time saved3326 sources35 min read