February 12, 2026 — A critical security race is defining the future of autonomous AI agents. Near.AI co-founder Illia Polosukhin is publicly building IronClaw, a Rust-based, security-hardened rival to the viral but vulnerable OpenClaw assistant. Concurrently, the AI protocol Olas has launched specialized autonomous agents, called Polystrat, to trade on the popular prediction platform Polymarket. These parallel developments, confirmed this week, signal a pivotal shift toward securing AI agents that handle sensitive credentials and automating complex financial strategies in decentralized markets.
IronClaw: A Security-First Rebuild of OpenClaw
Illia Polosukhin, a key figure behind the NEAR Protocol, loves OpenClaw’s capabilities but calls its current architecture a “total security black hole.” His response is IronClaw, an ambitious rebuild from the ground up in the memory-safe Rust programming language. The core innovation involves sandboxing each of the agent’s tools in isolated WebAssembly (Wasm) environments. This design means if one component is compromised or goes rogue, it cannot affect others or access the core system. Furthermore, IronClaw treats prompt injections—a common AI attack vector—as critical security risks and implements specific defenses.
The urgent need for this overhaul became starkly apparent this week. Polymarket traders experimenting with OpenClaw reported that the agent could still reveal private keys despite explicit user instructions forbidding it. “People are losing their funds and credentials using OpenClaw,” Polosukhin stated. “A number of people have stopped using it as they’re afraid it will leak all of their information. We started working on a security-focused version—IronClaw.” His solution is radical: the large language model (LLM) at the system’s heart will never directly touch secrets. Instead, credentials will reside in an encrypted vault, with the LLM granted temporary, permissioned access only for specific, pre-approved actions.
The Rapid Development and Inherent Risks of AI Agent Ecosystems
George Xian Zeng, General Manager of Near.AI, highlighted Polosukhin’s rapid execution. “He built the basis of it in one evening,” Zeng told Cointelegraph Magazine. “He was feeding his baby and building IronClaw at the same time.” This pace underscores the breakneck speed of innovation in the AI agent space. However, speed exacerbates existing risks. OpenClaw’s power stems from being a “harness” that controls multiple agents and tools with deep system integration, including terminal and browser access. This power, coupled with widespread JavaScript use, creates a large attack surface.
- Skill Marketplace Vulnerabilities: A critical unsolved problem is the security of third-party skills. Analytics firm SlowMist recently reported that 341 skills on the ClawHub marketplace contain malicious code designed to steal passwords or data. “The cool thing… is that anyone can build a skill. But the dangerous thing… is that anyone can build a skill,” Zeng acknowledged, suggesting a curated marketplace may be necessary.
- Near’s Interim Solution: While IronClaw develops, Near.AI Cloud offers a beta version of OpenClaw running in a Trusted Execution Environment (TEE), where all data is encrypted and inaccessible even to Near. This service also anonymizes requests to commercial AI models like Gemini and ChatGPT 5.2.
- The Competitive Landscape: Polosukhin’s critique implies users currently have limited, risky choices. His push for Rust leverages its memory safety to eliminate whole classes of bugs and its relative obscurity, which reduces the pool of hackers proficient in exploiting it.
Expert Insight: The Inevitability of Secure Agent Frameworks
David Minarsch, co-founder of Olas and CEO of Valory, whose company also deals with autonomous agent security, emphasizes architectural restraint. “That’s a key architectural design decision, which really restricts the capability of the agent,” Minarsch says regarding Olas’s approach to wallet security. “So, our fully structured agent won’t suddenly become your personal assistant. But it also means it’s safer.” This philosophy of limiting an agent’s scope to ensure safety mirrors the sandboxing approach of IronClaw, indicating a converging industry consensus on security-first design for powerful AI tools.
Olas Unleashes AI Traders on Polymarket
In a separate but related development, Olas has officially launched its Polystrat agents for Polymarket. These agents, adapted from the Omenstrat bots that have executed over 13 million transactions on the Gnosis-based Omen platform, represent a new wave of automated prediction market participants. Unlike simple arbitrage bots, Polystrat agents utilize a suite of tools—including news aggregators and public data feeds—to analyze and predict outcomes in markets that resolve within four days.
According to performance data shared by Olas, these agents have demonstrated a consistent edge in specific categories. Their success rate ranges from 59.2% to 63.6% in areas like sustainability, science, and business. However, performance drops significantly in subjective categories like fashion, arts, and social topics (37.96% to 48.57%), and is nearly neutral (51.01%) in sports betting. “What we see with Omenstrat is that over time they have a 55% to 65% success rate depending on which models and tools they use,” Minarsch explained, cautioning that Polymarket’s larger size and liquidity present a new test.
| Market Category | Agent Win Rate (Omen Data) | Notes |
|---|---|---|
| Science & Business | 59.2% – 63.6% | Strong performance on data-driven outcomes |
| Fashion & Arts | 37.96% – 48.57% | Poor performance on subjective topics |
| Sports | ~51.01% | Effectively neutral, no consistent edge |
| Overall (Omen) | 55% – 65% | Varies by model and tool configuration |
The Broader 2026 AI Landscape: Surveillance, Hype, and Automation
These technical advances occur within a contentious societal debate about AI’s role. Amazon faced immediate backlash after its Super Bowl ad for Ring doorbells showcased an AI “Search Party” feature, criticized for normalizing a default-on surveillance network. Meanwhile, the Super Bowl itself was saturated with ads from 16 different AI companies, leading to industry chatter about a potential market peak, reminiscent of the dot-com boom in 2000 and the crypto ad surge in 2022.
In media, tools like the text-to-video generator Seedance 2.0 are demonstrating rapid capability gains, while McKinsey reports AI is already streamlining film production by automating tasks like storyboard generation from scripts. This pervasive automation trend frames the developments at Near.AI and Olas not as isolated events, but as specific instances of a broader integration of AI into high-stakes digital environments, where security and reliability are paramount.
Market and Community Reactions
The crypto and AI developer communities have reacted with a mix of excitement and caution. Social media discussions highlight relief at IronClaw’s security focus but also concern over the speed of deployment. Polymarket traders are actively testing both OpenClaw configurations and the new Olas bots, sharing strategies and warnings in dedicated forums. The parallel launch of “attention markets” on Polymarket, in partnership with Kaito AI—where users bet on the virality and sentiment of topics—further illustrates the platform’s push toward AI-driven, meta-level prediction products.
Conclusion
The simultaneous emergence of IronClaw and Olas’s Polymarket bots marks February 2026 as a inflection point for autonomous AI agents. The narrative is bifurcating: one path focuses on hardening security for general-purpose assistants handling sensitive real-world tasks, while the other specializes agents for dominance in specific financial niches like prediction markets. Both paths, however, are guided by the same principle: for AI to be truly useful and trusted, it must be fundamentally secure and reliable. The coming weeks, as IronClaw moves from rapid development to public release and Polystrat agents face their first major test on Polymarket, will reveal whether these architectural solutions can meet the formidable challenges of a rapidly evolving automated landscape.
Frequently Asked Questions
Q1: What is the main security difference between OpenClaw and IronClaw?
IronClaw is being rebuilt from scratch in Rust, with each tool sandboxed in an isolated WebAssembly environment. Its core design prevents the large language model from directly accessing private keys or credentials, which are stored in a separate, encrypted vault. OpenClaw’s current JavaScript-based architecture has demonstrated vulnerabilities to credential leakage.
Q2: How do Olas’s Polystrat agents make predictions on Polymarket?
Polystrat agents do not simply hunt for arbitrage. They use a configured set of tools—including news feeds, data APIs, and analytical models—to assess the likely outcome of prediction market events that resolve within a short timeframe (under four days). Their strategy is based on information analysis, not market microstructure.
Q3: When will IronClaw be available to the public?
Near.AI executives indicate IronClaw is under active, rapid development, with a public release expected within “a matter of weeks.” In the interim, a more secure, cloud-based version of OpenClaw running in a Trusted Execution Environment is available through Near.AI Cloud’s beta program.
Q4: What are the biggest risks of using current AI agents like OpenClaw?
The primary risks are credential theft and financial loss. Agents require broad system access to function, and vulnerabilities in their code or malicious skills downloaded from marketplaces can compromise all connected accounts, cryptocurrency wallets, and private data.
Q5: How successful have AI agents been on prediction markets so far?
Data from Olas’s agents on the Omen platform shows a variable success rate between 55% and 65% over time, depending on the market category. They perform best on data-driven topics like business and science, and poorly on subjective topics like arts and fashion. Their performance on the larger Polymarket platform remains untested.
Q6: How does this affect the average cryptocurrency user or trader?
For traders, AI agents introduce new automated competitors and tools for market analysis. For all users, the security advancements pioneered by projects like IronClaw are crucial for safely adopting the next generation of AI-powered wallets, assistants, and DeFi tools that will manage digital assets.
