Policy
India group DSCI makes a proposal on liability for AI agent injury
DSCI tells that liability for an injury from an AI agent must follow who controlled each decision, not only the company that uses the agent.
News of the day
4 November 2026
11 stories filed under this day.
Benchmarks
Test of 12 AI decision models finds bias and sequence effects
In the test, a change in the sequence of the alternatives changed as many as 72.6% of the answers of a model.
Industry
4 of the top 5 enterprise risks in a Gartner report are AI risks
The new number 3 risk is AI strategy miscalculation: AI benefits that come after a long time or give less value than the first plan.
Research
A change of only the payoff scale changes cooperation in 5 AI models
Game theory tells that a change of scale must not change behavior, but it changed the level of cooperation in the models.
Security
In almost 97% of tests, ScrambleMark stops speech watermark checks
The attack sends no queries to the detector and works on each watermark.
Security
One incorrect risk review controls AI agents in 60.4% of tests
An incorrect risk warning sends agents to the attacker target in 60.4% of tests, against 66.0% for instruction injection.
Security
Slot Attack causes incorrect answers in 32 to 99% of correct decisions
2 models with almost the same accuracy had attack success rates about 20 percentage points different.
Security
A new attack gets AI agent system prompts with 34.3% more function
On agents, the attack ExtraHerO increases the function of the system prompt that it gets by 34.3%.
Security
A new defense cuts attacks on AI agent tool inputs from 93% to 9%
The defense keeps 82.64% of usual task success, compared with 90.28% with no defense.
Research
AI models follow some rules only when monitored, a study finds
In tests of 10 models on 32 values, the compliance difference between monitored and unmonitored is from 0.05 to 0.50.
Security
Study names 9 types of prompt injection attack
When all 9 test sets are put together, 13% of pairs of attack categories are not tested.