The Accountability Gap: The State of AI Agent Security and Governance 2026
SecurityHQ, Datadog · SecurityHQ und Datadog · 15 Oct 2026 · not peer-reviewed
Source finding
76% of 204 security leaders say responsibility for AI-agent decisions is formally assigned. Yet 45% report an incident in which responsibility was unclear.
49% say they can fully retrace an agent’s actions over the previous 30 days. After a critical incident, 17% say they could do so within minutes, 51% within hours, 23% within days, and 9% would take longer or are unsure.
Method and limitations
Survey of 204 security leaders at medium-sized and large businesses. These are self-reports, not observed incident tests. The public material does not explain country coverage, sampling or field dates. SecurityHQ and Datadog sell security and monitoring services. The survey partly groups agents and automation together. Evidence is exploratory.
Our interpretation
Test whether the responsible team can identify what the agent did, trace its permissions and stop it. A named owner alone is not proof of control.
Read first
The figures on reconstructing actions after a critical incident and on visibility over the previous 30 days.
Related evidence
- Control and accountability
- Control during real work
- Authority-Accountability Gap
- Agent Trace Recovery Time
- AI Accountability and Control Review
Open original study ↗
How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations
Daniel P. Gross, Dror Shvadron, Hansen Zhang · National Bureau of Economic Research · 1 Oct 2026 · not peer-reviewed
Source finding
The authors analyse more than 226,000 US STEM dissertation abstracts from January 2019 to May 2026. The Pangram detector finds almost no AI writing before 2023, compared with about 29% in the first five months of 2026.
Within the same programme and graduation cohort, graduates with detected AI text are less likely to enter academic research and more likely to move into industry or non-tenure-track roles. The association is strongest among US graduates.
Method and limitations
Dissertation abstracts are linked to later career information. The study is observational and has not been peer reviewed. Career plans may explain both AI use and job choice. A detector can misclassify text; abstracts do not reveal the full research or learning process. Public career profiles may be incomplete. Pangram supplied API credits and reduced prices. Evidence quality is moderate.
Our interpretation
Writing can be part of learning to think and judge. When AI takes over such work, plan alternative practice. This workplace implication is our inference: the study does not demonstrate that AI harms learning.
Read first
The abstract, the change from 2023 to 2026, the career analysis and the limitations of the AI detector.
Related evidence
- Learning from experience
- Protecting learning paths
- Learning that stays
- Learning Work
- AI Workforce Transition Review
Open original study ↗
