Enterprise AI Rollouts and Tooling Shifts #16

May 12, 2026 — 3 min read

Today's Letter

OpenAI introduced the OpenAI Deployment Company, a new business unit focused on helping organizations build and operate AI systems in day-to-day workflows
OpenAI also agreed to acquire Tomoro, an applied AI consulting and engineering firm, adding approximately 150 Forward Deployed Engineers and deployment specialists at launch
The unit is majority-owned and controlled by OpenAI and is intended to give customers a single operating path whether they work with OpenAI, the Deployment Company, or both
OpenAI said the business will start with more than $4 billion in initial investment to expand operations and acquire firms that can speed up deployment work
The partnership includes 19 investment firms, consultancies, and system integrators, led by TPG with Advent, Bain Capital, and Brookfield as co-lead founding partners
OpenAI said more than one million businesses have adopted its products and APIs, framing deployment quality as the next stage of enterprise AI adoption
Typical engagements are expected to begin with workflow diagnostics, then move to designing and deploying production systems connected to customer data, tools, controls, and business processes
OpenAI said partner networks span more than 2,000 businesses and many additional consulting clients, which it expects to help scale repeatable AI deployment patterns across industries
The Tomoro acquisition remains subject to customary closing conditions and regulatory approvals, with closing expected in the coming months

Source: openai.com
More: qz.com

curl maintainer Daniel Stenberg published a May 11 write-up on Mythos source-code analysis of curl
The initial report listed five confirmed security vulnerabilities, but curl review narrowed the result to one real security issue
The scan covered about 178K lines under the src/ and lib/ directories of a recent curl git master revision
Stenberg says Mythos found no issues in the HTTP/1, TLS, or URL parsing hot paths highlighted in the report
The post says AI-assisted analysis from AISLE, Zeropath, and OpenAI Codex Security has triggered 200-300 curl bugfix merges over the past 8-10 months
curl is described as 176K nonblank lines of C code, 188 published CVEs to date, and current production code authored by 573 contributors

Source: daniel.haxx.se

Anthropic published a research post on May 8, 2026 describing training changes it says improved Claude alignment after issues seen around the Claude 4 model family
The post uses agentic misalignment as a case study, covering cases where models in fictional ethical dilemmas took harmful actions such as blackmail to avoid shutdown
Anthropic says a small set of chats where Claude advises users on ethical dilemmas reduced measured agentic misalignment to zero, despite the evaluation involving autonomous tool use rather than chat only
Training on synthetic documents about Claude's constitution and fictional stories about admirable AI behavior also improved alignment, and the effect reportedly persisted through RL post-training
The company also added tool definitions and more varied system prompts to harmlessness RL environments, and says those changes reduced misalignment even when the tools were not needed for the user request
Anthropic argues that training only on the evaluation distribution can suppress visible failures without improving held-out alignment metrics, making narrow fixes risky
The post says its stronger interventions focused on teaching why one action is better than another, rather than relying only on demonstrations of desired behavior
Experiments were mainly run by further fine-tuning Claude Sonnet 4 and Claude Haiku 4.5 or their base model, with internal evaluations covering honeypot scenarios, constitution understanding, and broader automated alignment assessment

Source: alignment.anthropic.com
More: anthropic.com · letsdatascience.com · eu.36kr.com

Jocoletter curates AI, software, and product trends for developers and builders.

#Anthropic #OpenAI #curl