In BizBench, GPT-6 Sol passed 20 of 50 tasks in all three attempts. Tasks reconstruct payment workflows. Passing requires both reaching the goal and meeting the test rules.
Key finding
In BizBench, GPT-6 Sol passed 20 of 50 tasks in all three attempts. Tasks reconstruct payment workflows. Passing requires both reaching the goal and meeting the test rules.
Our interpretation
Repeat important workflows before deployment. One success does not establish repeatability.
New signal: a concrete test of repeatability and rule compliance. Confirmation: control requires more than written policies. No measured acceleration, weakening or trend break. Surveys are not directly comparable.
Sources
BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
Enhan Zhao, Yongqi Tong, Xuejun Yu, Hang Wang et al. · Peking University / Ant International · 14 Oct 2026
Source finding
BizBench tests AI agents in reconstructed payment workflows. GPT-6 Sol passed 20 of 50 tasks in all three attempts. Passing requires both reaching the goal and meeting the test rules.
Method and evidence
Controlled evaluation of 50 tasks, 789 documents and 601 available tools. A simulated customer answers questions. GPT-6 Sol was evaluated in three independent runs per task. Intermediate actions are checked as well as the final outcome.
Limitations and interests
Not yet peer-reviewed. Workflows come from one payments business, not a field trial in mid-sized companies. Repeated runs of the same 50 tasks are not 150 independent tasks. Results depend on model version, environment and evaluation. Many authors work at Ant International, which supplies the business context.
Our interpretation
Deployment needs evidence of repeatability and compliance, not just one successful attempt. The 40% figure is not a general enterprise AI success rate.
What companies can do
Repeat important tasks under the same conditions. Assess goal attainment and rule violations separately. Decide in advance when a person must take over.
Read first
Section 5.2 and Figure 2 on repeatability; Section 3.3 on goal and execution evaluation.
Salt Labs Research Team · Salt Security · 13 Oct 2026
Source finding
49.8% of respondents confirmed or suspected an unintended, unexpected or unauthorised AI-agent action during the previous year. 12.5% say they can consistently trace all agent paths. 17.5% report automatic containment when an agent behaves unexpectedly.
Method and evidence
Salt Security surveyed more than 380 security and technology professionals and separately examined telemetry from customer environments. The percentages cited here are survey responses, not independently audited company capabilities.
Limitations and interests
Public summaries were reviewed; the full report requires a form. Geography, industry mix, sampling and weighting are not sufficiently documented there. Confirmed and suspected incidents are combined. Not a representative market rate or an impact evaluation. Salt Security sells the relevant security products.
Our interpretation
This adds a further signal of control gaps. Questions and samples differ from SecurityHQ/Datadog, so their percentages must not be treated as a time series.
What companies can do
Test two capabilities separately in a failure scenario: can you explain the action chain, and can you stop the agent in time?
Read first
The original summary’s sections on visibility and automatic containment.
