Happy Thursday. The security bill for AI agents is coming due. A wave of acquisitions worth $96 billion in a single year shows that established vendors are racing to own the control points before the market sets.
Underneath that story, two more pieces of the agent stack are shifting. OpenAI's mobile voice rollout removes a friction point that has kept agentic tools tethered to desks, and new research out of Tsinghua attacks the latency cost of chaining models together, cutting one handoff step from 1,312 milliseconds to 90.
SECURITY & PRIVACY
AI Agent Security Draws $96B in M&A as Vendors Race for Control Points
Momentum Cyber counted $96 billion across 400 cybersecurity transactions in the past year, up 270% in deal value. Palo Alto Networks spent more than $900 million across two agent security acquisitions alone, and 35 of the 402 vendors in the space have already been bought, with acquisitions in 2026 running at nearly twice the 2025 pace.
The strategic logic here is specific, not broad. Buyers are not purchasing a category. They are acquiring control over discrete chokepoints: agent identity, prompt security, data classification, and traffic monitoring. If you are building software that grants agents access to enterprise systems, your security posture around non-human identities, API keys, service accounts, and auth tokens is the thing acquirers and auditors will scrutinize first. The vendors who own those controls will have significant pricing power as deployments scale.
ARTIFICIAL INTELLIGENCE
ChatGPT Work Tab Hits Mobile With Voice, Runs on New GPT-6 Models
OpenAI pushed its Work tab to iOS and Android on Sept. 23, letting Plus and Pro subscribers trigger multi-step workflows by voice. The update runs on three new GPT-6 models, including Sol and Luna, which launched the day before at roughly half the token cost of their GPT-5.6 equivalents.
The pricing on Sol and Luna matters for builders. At $2 per million input tokens and $10 per million output for Sol, and $0.10 and $0.50 for Luna, OpenAI is compressing the cost curve for agentic workloads while simultaneously expanding the user surface with voice on mobile. For software companies in ChatGPT's plugin ecosystem, mobile voice access is a new channel. Tasks that users previously had to sit down to complete can now start from anywhere, which changes the usage patterns worth designing for.
ARTIFICIAL INTELLIGENCE
Cache-to-Cache Protocol Cuts Model Handoff Latency 14x in Lab Tests
A team from Tsinghua University, Shanghai AI Lab, and Infinigence AI has built a method that lets AI models hand off through their internal memory states rather than generated text. In tests, replacing 80 text tokens with a direct cache transfer cut one relay step from 1,312 milliseconds to about 90 milliseconds, and accuracy improved 3.1 to 5.4 percentage points over text-based handoffs.
For teams building multi-model pipelines, the communication layer is the overlooked cost. Most architectures treat model handoffs as a fixed tax and optimize elsewhere. C2C reframes that assumption, though with a hard constraint: it requires direct access to model internals, so it cannot work with closed API providers unless those providers expose KV-cache states. The team plans to release a serving system this month. Watch whether the major inference providers move to support cache-level access, because that is the gate between a research result and a production primitive.
Created by the robots at The SaaS Sentinel