Agents Gain Power. Guardrails Get Serious. Robotaxis Scale.

What happened
AIR emerged from stealth with $50 million across two seed rounds to build security infrastructure for enterprise AI agents. Its platform discovers agents, continuously vets the skills, plugins, MCP servers, and other components they use, and can block interactions with unapproved software or external sources; AIR says it currently rejects roughly 27% of the publicly available add-ons and skills it evaluates.

Why it matters
The agent-security problem is shifting from securing the model itself to securing the software supply chain around the model. As agents gain permission to browse, load tools, access enterprise systems, and take actions, every external skill or MCP connection becomes another path through which bad instructions or compromised software can influence what the agent does.

What’s next
AIR says it already has more than 20 customers and is seeing its strongest demand in regulated industries including financial services and pharmaceuticals. Watch agent governance become a standalone enterprise infrastructure category as companies move from experimenting with agents to controlling what entire agent fleets are allowed to touch.

What happened
National-security startup Aslan publicly launched after raising $20.8 million, revealing agents designed to work inside underground criminal forums, Telegram channels, and other digital environments under human oversight. CEO Chase Reid told Axios that in one operation, Aslan agents ran autonomously for 21 days, with two entering a Telegram group involved in coordinating smuggling and extracting operational intelligence.

Why it matters
This is a significant jump from agents that complete office tasks or short computer workflows: Aslan is deploying agents into long-running, adversarial environments where they must maintain context, gather evidence, interact with people, and pursue an objective over days or weeks. The FBI, Homeland Security Investigations, and other agencies have already worked with the company on investigations, according to its CEO.

What’s next
Aslan plans to expand its technical and forward-deployed engineering teams as agencies test guardrails, data handling, and the boundaries of autonomous operations. The bigger test will be whether persistent agents can be governed tightly enough for sensitive government missions without eroding human accountability.

What happened
OpenAI said its forthcoming Astra model has reached the Critical cybersecurity capability threshold under its Preparedness Framework, the first OpenAI model to receive that designation. OpenAI says that with appropriate tools and access, Astra can identify previously unknown vulnerabilities and develop working exploit paths against well-protected systems without a person guiding every step.

Why it matters
The important shift is autonomy, not simply better coding. OpenAI's Critical threshold covers systems capable of developing zero-day exploits across hardened targets without human intervention or devising and executing an end-to-end cyber strategy from a high-level objective, putting frontier agent capabilities into territory where deployment controls become part of the product itself.

What’s next
OpenAI says Astra will be released soon, but its most advanced cybersecurity capabilities will initially be restricted to a small group of testers before defensive access expands through Daybreak Blue. Expect frontier-model launches to increasingly arrive with capability-specific access tiers, rather than one universal level of model access for every user.

What happened
Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 , pairing higher performance on coding and long-running work with a major reduction in cache-read pricing. Anthropic estimates Fable 5.1 costs about 25% less than Fable 5 on typical token-billed workloads and as much as 45% less on highly agentic workloads, while Mythos 5.1 remains restricted to vetted cybersecurity and life-sciences users.

Why it matters
For enterprises, the economics of persistent AI may matter as much as benchmark gains: tool-heavy agents repeatedly reuse large amounts of context, so cheaper cached context directly attacks one of the cost bottlenecks in long-running workflows. Anthropic is also preparing Enterprise Frontier Safeguards, which will keep protected data in cloud infrastructure controlled by the customer while preserving automated misuse controls.

What’s next
Fable 5.1 is available across Anthropic's platforms as well as AWS, Google Cloud, and Microsoft Azure, while Enterprise Frontier Safeguards are scheduled to begin rolling out later this fall. The competitive frontier is increasingly becoming a three-part equation - capability, cost per completed workflow, and enterprise control.

What happened
Cambridge University spinout Flower Labs launched Endeavor 1.0, its first production-ready Endeavor release, as a preview for select organizations. Flower describes it as a frontier class generalist built for reasoning, coding, and long horizon agent work, with the unusual option to run the model as a managed service or deploy it inside infrastructure controlled by the customer.

Why it matters
Most frontier intelligence still comes through vendor controlled APIs, creating tension for governments and enterprises that want stronger control over sensitive data, infrastructure, and model dependencies. Flower's bet is that organizations will increasingly demand high end AI that can become part of their own technology stack rather than permanently renting intelligence from a handful of closed providers; its benchmark competitiveness is company-reported and should still be independently validated.

What’s next
Flower is onboarding a limited group while increasing compute capacity before a wider rollout. Watch whether private deployment becomes a meaningful differentiator at the frontier especially in healthcare, finance, government, and other environments where sovereignty and data control can outweigh API convenience.

What happened
Waymo began welcoming its first public riders in Denver, San Diego, and Tampa on September 1, bringing the company's fully autonomous ride service to 14 cities. Waymo says tens of thousands of people in each of the three new markets have already expressed interest, with access expanding gradually toward broader public availability.

Why it matters
Physical AI ultimately has to survive outside benchmarks and controlled demos. Waymo's expansion shows one of the most mature embodied AI systems moving through the harder commercialization phase: adapting autonomous perception, planning, and control to multiple real cities while delivering an everyday consumer service without a human driver.

What’s next
Waymo says it will steadily add riders and expand service areas after months of local validation. The metric to watch is no longer simply whether autonomous driving works it is how quickly a proven autonomous system can replicate across new geographies without recreating the deployment process from scratch.

Agents are getting more capable, more persistent, and more exposed. Security is moving outward from the model to the tools and connections around it, frontier systems are crossing into higher-risk autonomous work, and enterprises are demanding stronger control over where intelligence runs. Meanwhile, physical AI keeps moving into everyday deployment. The next phase of AI won’t just be about what agents can do it will be about how safely, economically, and reliably we can let them keep doing it.

Compare a long-running agent with a tightly controlled one.

Take one real multi-step task like research, coding, analysis, or investigation and run it with a capable agent such as Claude Fable 5.1 or another tool that can work across a longer sequence of steps.

Give it:

Then run the same task again with broader permissions and fewer checkpoints.

Compare:

💡 Insight
Today’s AIR, Aslan, Astra, and Claude 5.1 stories all point to the same tension - better agents can operate longer and do more, but capability compounds faster than confidence unless permissions and oversight scale with it.