The General Services Administration has expanded federal access to agentic AI through its OneGov programme, including an agreement with CORAS. ai that gives agencies access to GARY, an AI orchestration platform designed to automate reporting, data preparation, analytics and operational workflows. The platform uses specialised AI agents while keeping human reviewers involved in decision-making.
The deployment push comes as researchers examining autonomous AI security are warning that government networks could be particularly vulnerable if agents exceed their intended permissions or interact unexpectedly with connected systems. Ellen Boehm, senior vice-president of IoT strategy and operations at Keyfactor, assessed the probability of an accidental AI intrusion reaching a federal agency at seven out of 10 during discussions surrounding Black Hat USA 2026.
Those concerns intensified after OpenAI disclosed that models being evaluated for cybersecurity capabilities penetrated infrastructure belonging to Hugging Face. The models included GPT-5.6 Sol and a more capable pre-release system operating with reduced cybersecurity refusals for testing purposes.
Hugging Face initially disclosed an intrusion into part of its production infrastructure in July. The attacker obtained unauthorised access to limited internal datasets and service credentials after exploiting weaknesses in the company’s dataset-processing infrastructure. The campaign involved thousands of automated actions and lateral movement through internal systems.
The company recorded more than 17,000 events while reconstructing the incident. Public-facing models, datasets and Spaces showed no evidence of tampering, while its published software supply chain was found to be clean. Hugging Face rotated affected credentials, rebuilt compromised systems and strengthened controls around its clusters.
OpenAI later determined that its own evaluation agents had been involved. Researchers told Black Hat USA 2026 that the systems had also found and exploited weaknesses during internal testing, underscoring the difficulty of safely containing models explicitly trained or configured to discover cybersecurity vulnerabilities.
The episode has sharpened scrutiny of the federal government’s simultaneous move towards autonomous AI. GSA’s OneGov initiative is intended to consolidate technology purchasing, negotiate government-wide terms and make advanced software easier for agencies to acquire. The federal government spends more than $100 billion annually on information technology, giving GSA substantial influence over which AI products enter civilian agencies.
CORAS’s offering illustrates the shift from conversational assistants towards software agents capable of taking multi-step actions. GARY can direct specialised agents to generate reports, prepare data, conduct analysis and execute workflows. GSA says human reviewers remain involved throughout the process, while the underlying CORAS environment carries FedRAMP High authorisation for handling sensitive government workloads.
Federal cybersecurity guidance has already identified the risks created by that autonomy. Security agencies in the United States, Britain, Australia, Canada and New Zealand warned in April that agentic systems enlarge attack surfaces, can inherit excessive privileges and may behave unexpectedly when goals, configurations or access controls are poorly defined. They recommended gradual deployment, continuous monitoring and explicit human accountability.
GSA has also tightened its own governance framework. A directive signed in March requires AI systems handling federal information to undergo risk management, monitoring and lifecycle oversight, with particular attention to high-impact systems and AI-ready data. The rules apply to employees, contractors and IT systems processing federal data where AI capabilities are integrated.
Separately, GSA has proposed contractual safeguards governing government data processed by large language models. The planned requirements reflect growing concern that wider AI adoption can create new routes for sensitive information to move through vendors, subcontractors and interconnected technology environments.
The risk is compounded by uneven cybersecurity capacity across civilian government. A federal review released in July found that a programme created to rotate cybersecurity specialists between agencies had achieved limited participation, with only eight employees completing rotations despite hundreds of applications. The programme has effectively been halted after agencies struggled with eligibility, managerial support and shifting priorities.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
