At Wednesday’s hearing of the Senate Homeland Security and Governmental Affairs Disaster Management subcommittee, Sen. Josh Hawley (R-Mo.) previewed legislation at odds with White House restraint: an upcoming bill would make AI firms liable for reckless design and users liable for reckless deployment, and would clarify that criminal hacking penalties apply to companies and users who deploy agents to commit crimes. Hawley invited Sam Altman; Altman declined (the invite went out Friday). OpenAI said it is deeply engaged with Congress.

Witnesses included Marius Hobbhahn of Apollo Research, who urged embedded evaluations and preserving chain-of-thought, and Paul Ohm of Georgetown, who pointed to FTC UDAP authority and torts and floated strict liability. Sen. Richard Blumenthal backs a related Hawley DOE adverse-incident testing bill. Other GOP senators stressed competition with China.

Sen. Elizabeth Warren and colleagues sent letters to the CEOs of Amazon, Google, Meta, and Microsoft seeking detail on tax breaks and lobbying around One Big Beautiful Bill Act incentives. The law includes 100% bonus depreciation for qualifying capital (for example AI data centers) and an updated R&D credit covering AI-related wages, hardware, and cloud training costs.

Reporting cites last-year depreciation breaks — Microsoft roughly $12 billion, Amazon $6.5 billion, Meta $4.9 billion, Alphabet more than $3.3 billion — and a roughly $96 billion lost federal revenue narrative tied partly to AI tax breaks, with corporate tax revenues down about 23%.

Kong Context Management is generally available in Kong AI Management as of September 30. The problem it targets: naive API-to-MCP conversion turns every operation into a tool, driving context tax, worse tool selection, token cost, and latency. Code Mode exposes a compact MCP interface — search the available API surface, retrieve schema on demand, then execute — so agents can compose multi-step operations in code (loops, conditionals, error handling) and cut round trips.

The product unifies APIs, existing MCP servers, and enterprise data sources via Kong AI Gateway–generated MCP interfaces. Tagline: agents access more of the enterprise without processing all of it at once.

On September 30, UiPath and Snowflake announced an expanded partnership: a two-way integration without copying data. UiPath Data Fabric reads and models Snowflake data in place; automations can also be triggered from Snowflake. UiPath Cartographer and Delegate plus Cortex connectivity support discovery in design and build. Snowflake CoCo can call the UiPath agentic skills library. The UiPath listing is live on Snowflake Marketplace.

Quotes from Snowflake’s Kasiviswanathan and UiPath’s Malpani emphasize zero-copy grounding and a Map of Work that keeps governed data connected to automation.

As of September 24 Public Preview, Google Cloud API Gateway can act as a remote MCP server: annotate OpenAPI 3.x, deploy, and operations become MCP tools. MCP JSON-RPC lands on /mcp; tools/call is transcoded to REST through the existing auth, quota, and logging path. Google positions it as a lightweight on-ramp (Cloud Run and similar); Apigee for full lifecycle; Agent Gateway for outbound agent call governance.

Limits to know: no empty-body 204 operations, nested schema limits, at most 1,000 tools per gateway, and MCP and model routing are mutually exclusive in the same config. Roadmap items include resources and prompts, streaming, and Model Armor.