AI Work & Model Security
OpenAI Releases ChatGPT Work Data as ARC Evals Tests Autonomous Agents
OpenAI published OpenAI Signals data covering 165 work activities across nearly 100 countries, ARC Evals evaluated four language-model agents across 12 autonomous replication tasks, and cybersecurity researchers analyzed the CVE-2026-111497 Brotli decompression flaw.

Your listening Queue & catch up
Add a story or build a catch-up queue to start listening.
Global workplace adoption of generative artificial intelligence, empirical benchmarks for autonomous language-model agents, and open-source library security disclosures converged as OpenAI released multi-country occupational usage data, ARC Evals published its autonomous task evaluation report, and enterprise defenders assessed a critical flaw in the Google-developed Brotli decompression library.
OpenAI Signals Maps 165 Work Activities Across Nearly 100 Countries
OpenAI published a new OpenAI Signals dataset examining how consumers use ChatGPT for work across 165 Intermediate Work Activities (IWAs) and nearly 100 countries, according to OpenAI News. Drawing on the standardized O*NET occupational framework, the quarterly dataset spans from Q1 2025 through Q3 2026 and covers work-related messages on individual ChatGPT plans, excluding enterprise accounts and Codex usage. The categorized tasks range from preparing instructional materials and developing marketing content to programming and analyzing business data.
Across 74 countries observed in every quarter of the study, the 10 most common work activities accounted for an average of 51.1% of each country's classified work messages in Q3 2026, down from 56.0% in Q1 2025, indicating that work-related ChatGPT usage became less concentrated over time as countries specialized in distinct occupational tasks. When comparing ChatGPT usage against actual employment-weighted tasks across U.S. jobs, researchers noted that about 43% of the employment-weighted task mix for mapped U.S. occupations falls into Other IWAs outside the individually reported categories, such as moving materials, operating vehicles, and maintaining health or medical records. To protect privacy, automated classifiers process messages after identifying details are removed, and OpenAI publishes aggregated statistics with statistical noise.
ARC Evals Tests 4 Language-Model Agents Across 12 Autonomous Tasks
In model safety research, ARC Evals released its first public evaluation report introducing a methodology to test whether large language model agents can acquire resources, create copies of themselves, and adapt to novel challenges—a capability set termed autonomous replication and adaptation (ARA), METR reported. Researchers operationalized 12 real-world tasks of increasing difficulty and built 4 example agents combining command-line scaffolding programs with Anthropic's Claude and OpenAI's GPT-4.
The 12 benchmark tasks ranged from searching a filesystem for a password, listing the top 10 most read BBC articles, and finding valid vote counts from the 2022 Austrian presidential election on Wikipedia to setting up a Bitcoin wallet, making a phone call through a Twilio API account, counting dogs in an image file, identifying two new employees who joined a company in the last 6 months, configuring the open-source GPT-J model on an Amazon EC2 server, writing an OpenAI API agent with an automatic VM restart script, conducting a simulated university phishing exercise under human oversight, and fine-tuning LLaMA-7B on AWS to extend context length. The Claude and GPT-4 agents completed only the easiest tasks, though researchers found that GPT-4-launch completed three additional tasks compared with GPT-4-early, demonstrating that general helpfulness fine-tuning can increase ARA-relevant capabilities.
Cybersecurity Experts Temper Alarms Over CVE-2026-111497 Brotli Flaw
Meanwhile, cybersecurity researchers urged organizations to prioritize practical risk assessment over panic following the disclosure of CVE-2026-111497, a critical buffer overflow defect in the Google-developed Brotli decompression library, according to CyberScoop. Discovered by OpenAI Codex Security and disclosed by Node.js in an Oct. 21 security advisory, the vulnerability—dubbed "Node2Shell" or "Broccoli2Shell"—can trigger system crashes and remote code execution across software packages and cloud services, prompting emergency patches from vendors including Microsoft and Python.
Wiz researchers warned that 91% of cloud environments have resources vulnerable to CVE-2026-111497 and 82% of cloud environments have publicly exposed Node.js instances, though researchers have not observed active exploitation in real networks. Jake Knott, head of threat intelligence at watchTowr, noted that real-world exploitability depends on how the library is implemented within specific software, while Christiaan Beek, vice president at Rapid7 Intelligence, and Caitlin Condon, vice president of research at VulnCheck, cautioned that a surge of AI-discovered vulnerability disclosures is straining security teams with constant emergency alerts.
Sources
Comments (0)
No comments yet. Be the first to share a thought on this story.
Nominate a neighbor or local businessRequest a correctionContact the newsroom