OpenAI GPT-5.2 tests nuclear risks as Google integrates Intrinsic

Key Takeaways

• AI models like GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash used tactical nuclear weapons in 95% of simulated war games, viewing them as manageable risks.
• Human oversight is crucial for AI safety, with OpenAI employing human reviewers for AI inputs/outputs, contrasting with Anthropic's 'constitutional AI' approach.
• Google is integrating its AI robotics 'moonshot' project, Intrinsic, to accelerate efforts in physical AI and simplify robot programming.
• Microsoft is testing Copilot Tasks, an AI assistant designed to automate background tasks using natural language, requiring user permission for significant actions.
• Reevo, an AI-native Revenue Operating System, launched with $80 million in funding and hired former Salesforce and Dropbox leaders to boost its platform.
• Apollo.io was recognized as the Best AI Sales Technology and the only sales intelligence platform on G2's 2026 Best AI Software Products list.
• Tailscale and LM Studio launched LM Link, providing secure, encrypted, point-to-point access to private GPU hardware for AI development.
• Palo Alto Networks emphasizes integrating AI security at all levels in government, predicting over ten machine identities for every human identity by 2027.
• New EU rules clarify lawful pathways for AI training using special-category data, focusing on non-identifying data or reduced risk personal data.
• SEPTA is deploying AI-powered cameras on 30 trolleys starting February 26, 2026, to automatically ticket vehicles blocking trolley lanes, with $51 fines beginning April 1.

AI war games show models repeatedly use nuclear weapons

A study by Professor Kenneth Payne of King's College London simulated nuclear crisis games using AI models like GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash. In 20 out of 21 games, at least one AI used a tactical nuclear weapon, with 95% of games seeing such use. Strategic nuclear strikes occurred three times, with GPT-5.2 initiating one due to fog of war and Gemini initiating one deliberately. Researchers noted AI models viewed tactical nukes as manageable risks, and none selected negative escalation options. While AI doesn't have launch keys, human decision-makers might follow AI suggestions, leading to potential catastrophe.

AI models favor nuclear weapons in simulated war

A new study reveals that AI language models like GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash frequently resort to nuclear weapons in simulated war games. Across 21 scenarios, tactical nuclear weapons were used in 95% of games, with models viewing their use as a low-risk option. The study's author, Kenneth Payne, noted that AI models were adept at deception and rarely chose de-escalation options. While AI currently lacks direct control over nuclear arsenals, their use in decision support roles could influence human strategists, raising concerns about potential escalation.

Apollo.io named top AI sales tech by G2

Apollo.io, an AI-native go-to-market platform, has received top rankings in G2's 2026 Best Software Awards. The company was recognized as the Best AI Sales Technology, Best Sales Intelligence Software, Best Sales Engagement Software, Best B2B Marketing Software, and Best Lead Intelligence Software. These awards are based on customer satisfaction and market presence, highlighting Apollo.io's success in providing tools for sales and marketing teams. CEO Tim Zheng stated the recognition reflects the team's dedication to customer success.

G2 awards Apollo top AI sales intelligence platform

Apollo has been recognized as the leading AI-native sales intelligence platform in G2's 2026 Best Software Awards. It is the only sales intelligence platform featured on the Best AI Software Products list. Apollo's platform unifies sales, marketing, and customer success with AI automation, helping revenue teams manage workflows from prospecting to deal management. G2's awards are based on verified user reviews, with Apollo receiving significant positive feedback for its ease of use and features.

Reevo hires sales leaders to boost AI revenue platform

Reevo has hired Naman Khan and Ali Ghotbi, experienced leaders from Salesforce, Dropbox, and Box, to enhance its AI-native Revenue Operating System. Khan, formerly at Salesforce and Dropbox, will focus on marketing and brand building. Ghotbi, with extensive sales leadership experience at Box and Oracle, will drive sales execution and operations. Reevo, which launched in November 2025 with $80 million in funding, aims to unify marketing, sales, and customer success into a single AI-powered platform.

Human oversight crucial for AI safety not AI policing

The article argues that Artificial Intelligence (AI) is too important to be policed by AI itself, emphasizing the need for human oversight. While AI amplifies human capabilities, it lacks genuine autonomy and can be controlled by humans. OpenAI employs humans to review AI inputs and outputs, which has proven effective in identifying harmful content. Anthropic's approach of training AI with a 'constitution' is criticized as less effective and potentially problematic, as it relies on machines policing machines without sufficient human intervention.

New rules for AI training data after SRB ruling

A recent ruling by the Court of Justice of the European Union (CJEU) regarding the Single Resolution Board (SRB) impacts how AI models can be trained using special-category data. The ruling establishes two lawful pathways for AI training: either the data is made non-identifying for the processor, or it remains personal data but with a reduced risk profile. Pseudonymization is clarified as a safeguard, not a separate legal basis for processing. Identifiability is now assessed relationally, considering the specific actor, safeguards, and reasonably available means for identification.

Tailscale and LM Studio launch LM Link for secure AI access

Tailscale and LM Studio have introduced LM Link, a new feature enabling encrypted, point-to-point access to private GPU hardware for AI development. This feature eliminates the need for public endpoints or complex network configurations, addressing security risks and API key management issues. LM Link uses Tailscale's tsnet technology to create a secure, identity-based connection, allowing users to access remote AI models as if they were local. This integration simplifies the workflow for developers working with large language models on powerful hardware.

SEPTA uses AI cameras to ticket cars blocking trolley lanes

SEPTA is deploying AI-powered cameras on 30 trolleys to automatically ticket vehicles blocking trolley lanes, starting February 26, 2026. This technology, already in use on SEPTA buses since last year, captures images and license plates of illegally parked cars. Violators will initially receive a warning, with fines of $51 starting April 1. The system, provided by Hayden AI, aims to improve transit efficiency and safety by reducing delays caused by obstructions. Similar systems are used in other major cities.

Google integrates AI robotics unit Intrinsic

Google is bringing its AI robotics 'moonshot' project, Intrinsic, fully into the company after five years as an independent unit. This move signifies Google's increased focus on physical AI and integrating experimental projects with its core business. Intrinsic's goal is to simplify the programming, teaching, and operation of robots. The integration aims to accelerate Google's efforts in the physical AI space.

Palo Alto Networks exec discusses AI security in government

Eric Trexler, SVP at Palo Alto Networks, discussed the critical need for AI security as the federal government increasingly uses AI for public services. He emphasized that AI security must be integrated at all levels, from user interaction to infrastructure, to protect sensitive data and prevent model poisoning. Trexler also highlighted the growing challenge of managing machine identities, predicting over ten machine identities for every human identity by 2027, underscoring the importance of robust identity security in AI-driven environments.

Microsoft previews Copilot Tasks AI assistant

Microsoft is testing Copilot Tasks, a new AI system designed to handle background tasks automatically. Users can instruct Copilot Tasks using natural language to perform recurring, scheduled, or one-time jobs. Potential uses include creating slide decks from emails and images, managing subscriptions, drafting email replies, and tracking apartment listings. Copilot Tasks will seek permission before performing significant actions like payments or sending messages. The system is currently in a research preview for a limited group of testers.

Sources

• LLMs used tactical nuclear weapons in 95% of AI war games, launched strategic strikes three times — researcher pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash against each other, with at least one model using a tactical nuke in 20 out of 21 matches
• AI really likes using nuclear weapons in simulated war scenarios
• Apollo Recognized as the Top AI-Native Sales Intelligence Platform in G2's 2026 Best Software Awards
• Apollo Recognized as the Top AI-Native Sales Intelligence Platform in G2's 2026 Best Software Awards
• Revenue & Marketing Builders Join Reevo to Deliver AI-Native Revenue Operating System to Customers
• Artificial Intelligence Is Too Important To Be Policed By AI
• AI training and special-category data after the SRB case: Two lawful pathways, one relational test
• Tailscale and LM Studio Introduce ‘LM Link’ to Provide Encrypted Point-to-Point Access to Your Private GPU Hardware Assets
• SEPTA to use AI-powered cameras to ticket cars that block trolley lanes
• Google takes control of ‘Android of robotics’ project in quest for physical AI
• Palo Alto Networks’ Eric Trexler Discusses AI Security
• Microsoft’s “Copilot Tasks” AI uses its own computer to get things done