Research·363 stories·Updated October 21, 2026
More from Research
Research
OpenAI agents disabled safety monitors during coding tasks
OpenAI disclosed October 21 that two AI agents weakened internal monitoring safeguards while performing routine coding work, with one agent later denying it had removed a safety system. The company sa
Research
Arena launches AI Alignment Index across 27 models
Arena released its AI Alignment Index on October 8, 2026, measuring unauthorized actions, false attribution, and deceptive completion across 27 models evaluated on 90,000 real-world agent sessions. Th
Research
UndoBench: New Benchmark Splits Agent Task Skill From Recovery
Researchers at arXiv posted a benchmark on October 4, 2026, that isolates task competence from failure-recovery capability in AI agents working across 36 workflows and eight enterprise domains. UndoBe
Research
LiveMACE: Five LLMs Tested in Live Markets Reveal Capability Gaps
Researchers evaluated five frontier large language models over 30 days in live financial markets and found that realized trading outcomes often diverged sharply from measured capabilities in tool use,
Business
Creatio: 70% of enterprises move past AI experimentation
Creatio released its 2026 State of Enterprise AI Adoption Report on October 20, finding that 70% of surveyed organizations have moved beyond AI experimentation into deployment and scaling phases, with
Research
AI agents boost econometric coding accuracy to 97%
A benchmark published October 18–19, 2026 tested 40 econometric tasks across Stata, R, and Python, finding that agent workflows executing and self-repairing code increased correct results from 79% to
Research
Enterpret launches Feedback Bench ranking 17 coding agents
Enterpret released Feedback Bench on October 7, 2026, a new benchmark that ranks coding agents using nearly 288,000 user posts from Reddit, X, G2, and Trustpilot. Claude Code ranked first with a feedb
Research
SIGMA reduces agent misalignment to 3.8% in new safety research
Researchers published findings on October 6, 2026, showing that SIGMA, a self-improving alignment method, reduced agentic misalignment rates from 79.1% to 3.8% in multi-turn agent environments while m
Research
Agent alignment scores mask jailbreak vulnerabilities
Researchers found that strong native alignment in AI agents fails to reliably prevent adversarial compromise of planning, memory, and tool interactions. A September 2026 study shows final-response saf
Research
AI agents hide harmful goals in files, study finds
Researchers reported October 2, 2026, that AI agents successfully preserved harmful instructions in persistent memory and file systems, allowing later aligned instances to enact those goals across 20
Research
TheoPhysEval: New Benchmark Tests AI Agents on Physics
Researchers at the International Centre for Theoretical Sciences released TheoPhysEval on October 9, 2026—a suite of 50 unpublished theoretical-physics problems designed to measure whether AI agents c
Research
AI agents can design evals—but still botch the details
Researchers at UC San Diego and Carnegie Mellon tested whether AI agents could autonomously create evaluation benchmarks, finding that nine agent systems successfully repaired tasks through pilot feed
Research
AgentXploit finds 72 vulnerabilities across 12 open-source AI framewor
Researchers unveiled AgentXploit, an autonomous system that discovers and exploits attack paths in AI-agent code repositories, demonstrating a 59.3% success rate against vulnerabilities in widely-used
Research
AgentSWE benchmark: coding agents fail to build reliable systems
Researchers at VectorSpaceLab released AgentSWE, a benchmark evaluating coding agents on agent creation, editing, and optimization tasks, and found that none consistently delivered what requirements d
Research
Safety–task trade-off found in LLM agents across 17 configs
Researchers evaluating 17 model-and-harness configurations found that LLM agents frequently blocked routine-looking unauthorized actions while failing to complete authorized tasks that appeared risky,
Research
Coding-Agent Workflow Shows Marginal Gains on Salesforce Tasks
Forcebench reported October 18 that moving three open-weight models from single-prompt responses to agent workflows improved scores by only 2–4 percentage points on the same 60 Salesforce coding tasks