Research
AI agents that do tasks correctly leak more private data, a study finds
Leakage increases by a maximum of 3 times from 1 turn to 6 turns.
arxiv.orgClaimed, not confirmed
News category
Research
129 stories, newest first.
Research
The FPA method gives more correct answers on 190 memory questions
With DeepSeek Flash the score went from 0.825 to 0.920, and with Kimi it went from 0.782 to 0.916.
Research
7 AI agents write reports that agree with pseudoscience
No agent can resist the incorrect claims in more than 27.4% of the tests.
arxiv.orgClaimed, not confirmed
Research
An 8B model finds 2.89 disagreements that persons do not speak about
In 18 meetings, the persons did not speak about the disagreements that IoA-Prober-8B found.
arxiv.orgClaimed, not confirmed
Research
Attacks by many agents go around monitors that look at one agent
A planner agent increases the rate of completed attacks by a maximum of seven times.
arxiv.orgClaimed, not confirmed
Research
Claude Opus 4.8 follows 70.1% of coding rules in a new test
In this test, agents most frequently do not follow the rules for the sequence of work, tools, and reports.
arxiv.orgClaimed, not confirmed
Research
One incorrect document cuts research agent results by 66 to 88 points
In the DRNOISE test, agents find the correct records, but they use the document that gives a different answer.
arxiv.orgClaimed, not confirmed
Research
SwaT sends KV-cache data between LLMs of different families
In tests, SwaT had 40% more mean accuracy and was 4 times as fast as text.
arxiv.orgClaimed, not confirmed
Research
3 of 6 deepfake detectors flag almost all frames of a robot face
The authors write that the detectors flag a face that is not a human face. They do not find a forgery. Light JPEG compression makes the effect small.
Research
An incorrect preference makes AI models give incorrect medical advice
The errors add up with each turn of the chat, also when the model first gives the correct result.
arxiv.orgClaimed, not confirmed
Research
An open-weight model gets the same score as closed models at a low cost
The authors write that this is correct only when the model has sufficient turns. The limit on messages gives incorrect low scores.
Research
AI agent traffic on Mintlify sites is 2 times the human traffic
Mintlify says 83% of the companies in its survey let agents write updates to documentation. Only 9% let agents publish without human inspection.