Pentagon Shifts Classified AI to Secure Environments

The United States Department of Defense is undergoing a significant transition in its approach to generative artificial intelligence by migrating all classified workloads away from commercial instances of Anthropic’s models. This initiative, mandated to be completed by October, represents a strategic pivot toward hosting sensitive AI operations exclusively within highly secured, air-gapped environments rather than relying on public cloud infrastructures.

Imagen generada con IA

This decision underscores the military’s ongoing struggle to balance the integration of cutting-edge commercial large language models with the stringent security protocols required for national defense. While the Department of Defense has previously experimented with utilizing Anthropic’s Claude models through specialized software environments, the move signals a broader institutional preference for internal control. By restricting these specific classified functions to controlled, private environments, defense officials aim to mitigate risks associated with data leakage, unauthorized access, and supply-chain vulnerabilities inherent in broader commercial AI deployment.

For the technical community, this development highlights the growing divide between general-purpose AI development and the rigorous, specialized demands of defense sector computing. It suggests that while the DoD remains enthusiastic about leveraging the reasoning capabilities of advanced models, they are increasingly skeptical of the current security posture offered by external cloud providers for their most sensitive intelligence tasks. The transition is expected to push defense contractors and software teams to focus more heavily on local model deployment, containerized orchestration in disconnected environments, and the hardening of open-source or proprietary models that can operate without a persistent connection to commercial API endpoints.

Ultimately, this shift serves as a case study for organizations handling top-secret data that wish to integrate machine learning. The Defense Department is demonstrating that for high-stakes environments, the speed and accessibility of commercial AI must be secondary to architectural sovereignty. As the October deadline approaches, all eyes will be on whether the DoD can maintain the functional performance of these AI tools while operating entirely within their own internal computational infrastructure, effectively setting a new standard for secure AI adoption in the public sector.

Artículos relacionados de LaRebelión:

• IAs Dark Side Anthropic Reveals Weaponised AI Models
• Anthropic Researcher Resigns Over AI Safety Concerns
• Google, Anthropic y OpenAI lanzan nuevos modelos y programas de IA para ciberseguridad
• AI Cyber Defence Google OpenAI Anthropic Lead Charge
• Anthropic presenta Fable y Mythos 5.1 con recortes de precios

---

Fuente Original: DefenseScoop

Artículo generado mediante AI.larebelion.

IAs Dark Side Anthropic Reveals Weaponised AI Models

A recent report from Anthropic sheds light on the alarming ways malicious actors are exploiting AI models for dangerous purposes. The 'Misuse AI' report, detailing over 150 pages of documented cases, highlights how 'Generative Threat Groups' (GTGs) are leveraging these advanced technologies. These groups, encompassing criminal organisations, nation-states, activists, and crimeware firms, are moving beyond traditional cyberattacks.

The report categorises threats, starting with 'Generative Threat Groups' (GTG) which use AI for sophisticated cyber operations, including reconnaissance , malware creation, data exfiltration, and lateral movement. It's now evident that the landscape of pentesting and cybercrime has fundamentally shifted, with AI becoming an indispensable tool for both offensive and defensive operations. Beyond cyber warfare, the report delves into the concerning use of AI for widespread disinformation campaigns, mirroring techniques for politically motivated content manipulation and amplification through social media bots, as previously demonstrated in presentations.

Perhaps the most alarming revelations concern the use of AI in developing conventional and advanced weaponry, including biological weapons. Examples cited include the construction of missile guidance systems using mobile phones and the development of coordinated drone swarms for military attacks. Furthermore, the report details the creation of directed-energy weapons and the alarming prospect of AI being used to engineer biological agents for malicious purposes. This underscores t he critical need for robust security layers and ongoing research into AI alignment, especially as we approach the era of Artificial General Intelligence (AGI), where AI's self-improvement capabilities raise profound questions about control and safety.

Fuente Original: http://www.elladodelmal.com/2026/09/anthropic-explica-como-sus-modelos-de.html

Artículos relacionados de LaRebelión:

• Anthropic Researcher Resigns Over AI Safety Concerns
• DeepSeek Shakes Up Open Source AI Models
• Google, Anthropic y OpenAI lanzan nuevos modelos y programas de IA para ciberseguridad
• AI Cyber Defence Google OpenAI Anthropic Lead Charge
• Anthropic presenta Fable y Mythos 5.1 con recortes de precios

Artículo generado mediante LaRebelionBOT

Rethinking Enterprise Testing for Agentic AI

The emergence of agentic AI represents a fundamental shift in software architecture that renders traditional testing methodologies obsolete. Unlike deterministic software, where specific inputs yield predictable outputs, agentic AI operates with a degree of autonomy and non-deterministic decision-making. These systems function as dynamic loops that perceive their environment, reason through complex tasks, and execute multi-step workflows, creating a testing surface that is exponentially more variable and harder to validate than standard monolithic or microservices applications.

Imagen generada con IA

For enterprise software engineers, this evolution demands a transition from static test scripts to continuous, behavior-based evaluation frameworks. Traditional unit and integration tests are insufficient when the system’s logic evolves based on context and external feedback loops. Instead, organizations must implement robust guardrails and observability platforms that monitor agent performance in real-time. This includes measuring the reliability of reasoning chains, verifying the safety of external tool execution, and assessing the semantic accuracy of autonomous outcomes. Testing can no longer be viewed as a final pre-deployment phase; it must be embedded directly into the agent’s lifecycle to manage the inherent risks of drift and hallucination.

The core challenge lies in the unpredictable nature of how agents might interpret instructions or navigate unforeseen edge cases in production. To mitigate these risks, enterprises are moving toward sandbox-based simulations where agents are subjected to high-fidelity environments that mirror their operational reality. By using synthetic data and adversarial stress testing, engineers can identify vulnerabilities in an agent’s decision-making process before it is granted access to sensitive enterprise systems. This shift requires a cultural move toward rigorous monitoring, where the focus moves from confirming code functionality to ensuring the safety, ethical alignment, and task accuracy of the agentic workflow.

Ultimately, the move toward agentic AI necessitates a unified strategy for automated quality assurance. Developers must treat agents as evolving personas that require ongoing supervision rather than set-and-forget software components. As these autonomous tools become integral to enterprise automation, the ability to validate non-deterministic outcomes through automated evaluation pipelines will determine the success and scalability of AI-driven transformation within the modern stack.

Artículos relacionados de LaRebelión:

• Slopsquatting AIs New Software Supply Chain Threat
• SUSE Sale EQT Eyes 6 Billion Enterprise Software Deal
• El nuevo rol del ingeniero de software
• Ingenieros de Software Disenando Limites No Solo Codigo
• Securing Enterprise AI with Data and Boundaries

---

Fuente Original: IT Pro

Artículo generado mediante AI.larebelion.