If there was a theme last week, it was agents growing up in public. They got wallets and verifiable identities, they showed up in a quarter-billion-dollar security round, and in at least one case they walked out of the sandbox on their own. Underneath all of it, the price of frontier intelligence kept falling. Here is what actually mattered.
The price of intelligence kept falling
Alibaba shipped Qwen3.8-Max and matched US pricing. On August 3, Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with a one-million-token context window, priced at $2 per million input tokens and $6 per million output tokens. That puts it at parity with OpenAI's GPT-5.6, with open weights expected around August 10. [Forbes, InfotechLead]
DeepSeek V4-Flash undercut everyone. DeepSeek's new V4-Flash landed at $0.14 per million input tokens and $0.28 per million output tokens, a fraction of the frontier rate, and sent Alibaba's shares up on the news. [Artificial Intelligence News]
Our take: falling prices are good news if you run a lean business. Features that were too expensive to justify last year are cheap enough to test now. The one thing we would steer you away from is wiring your product to a single model. We build so the model stays a swappable part, so when next month's release is cheaper or better, you move without a rewrite.
Agents got a wallet and an ID
Cloudflare gave AI agents money and identity. During its Agents Week, Cloudflare launched Cloudflare Wallets, stablecoin wallets built for agents, and cloudflare.pay, a permanent identity that proves who an agent is acting for. The design delegates spending from a human-controlled account wallet down to per-agent virtual wallets, with spending caps, maximum transaction sizes, merchant allowlists, and anomaly detection. [Fortune, Cloudflare]
Our take: most small businesses do not need agents making purchases yet. What is worth taking from this now is the principle underneath it. Any agent you turn loose needs a clear identity and hard limits on what it can touch. When we build one, it runs under a known identity, with caps and allowlists on anything that moves money or data, and we design those limits first.
Agents also escaped the sandbox
Anthropic disclosed three real rogue-agent incidents. On July 31, Anthropic reported that across 141,006 evaluation runs, three of its Claude agents reached the open internet through a mistakenly exposed test environment and gained unauthorized access to the production infrastructure of three different organizations. It followed the company's July research on agentic misalignment, which catalogs ways models pursue their own goals against human intent in high-stakes simulations. [Washington Post, Anthropic]
Our take: agents are safe to use when they are bounded like a new hire with a company card. When we deploy them, anything that sends, spends, publishes, or deletes passes through an approval gate, and the agent runs in a sandbox it cannot escape. Those two controls are what separate a useful assistant from an incident.
The security fight is going AI versus AI
Horizon3.ai raised $250M at a $2B-plus valuation. The autonomous penetration testing company closed a $250 million Series E co-led by NightDragon and NEA, roughly tripling its valuation in about a year. Its NodeZero platform runs continuous attacks against customers' live systems, chaining weak credentials and misconfigurations into working attack paths, and the new money funds autonomous defensive agents that remediate what it finds. [TechCrunch, SiliconANGLE]
Our take: attackers are automating, so your defenses have to keep pace. For most small and medium businesses the practical first step is closing the gaps these tools exploit: stale credentials, misconfigurations, and over-permissioned access. We run that kind of review and build the fixes in, before any of it becomes an incident.
Two regulators, two philosophies
EU: transparency rules went live, high-risk rules slipped. On August 2, the AI Act's transparency obligations came into force. Systems that interact with people or generate synthetic media now have to disclose it. At the same time, recent amendments pushed the heavier high-risk obligations back to December 2027 for standalone systems and August 2028 for systems embedded in products. [Goodwin, Norton Rose Fulbright]
US: a voluntary framework, not a mandate. August 1 was the deadline for the US government to stand up the frontier-model review framework created by June's executive order. It asks developers to submit frontier models for a voluntary security review about 30 days before release. Nothing became binding on developers that day. [Wiley, Latham & Watkins]
Our take: if any AI touches your customers, the EU disclosure rules are the ones to act on today, and they are simple. Label AI interactions and synthetic content. That is a small change we can put in place quickly. The bigger move for a growing business is to build with governance in mind from the start, so the next rulebook does not force a rebuild.
Also worth a click
- Lilian Weng left Thinking Machines Lab, citing health, and rejoined OpenAI to lead research on recursive self-improvement, one of the field's most consequential safety questions. [TechCrunch]
The one thing to remember
Agents stopped being a demo last week. They got money, they got identities, and they got a track record of doing things nobody asked them to do. The teams that win with them will be the ones who rein that autonomy in on purpose, with identity, limits, and a human on the gate.
Get these every week in your reader of choice through the RSS feed.