6 min read · 12 sources

OpenAI is pushing deeper into automated scientific workflows by acquiring biopharma orchestration platform Orchestra Bio. The acquisition pulls Orchestra’s domain-specific pipeline tooling directly into Rosalind, OpenAI’s internal life sciences research initiative.

The goal is building agentic systems capable of tracking experimental wet-lab discoveries through clinical trial setups and regulatory paperwork. Biopharma engineering stacks have historically been a mess of bespoke spreadsheets, siloed laboratory information management systems (LIMS), and legacy document vaults. Automating this connective tissue marks a clear shift for OpenAI from general-purpose API provider to vertical operator in high-margin, heavily regulated industries.

Meanwhile, enterprise consolidation is accelerating, AI-assisted bug-hunting is taking PHP sites offline, and automated code generation is pushing build pipelines to their limits.

Databricks’ Agent Memory cuts token consumption by 31 % while tripling task success rates.

Salesforce drops $5.66 billion on Braze to anchor Claudeforce

Source: cnbc.com ↗

Salesforce has agreed to acquire customer engagement platform Braze for $5.66 billion in an all-cash transaction valued at $47.50 per share. The acquisition is scheduled to close in fiscal 2028. It follows Salesforce’s $3.6 billion acquisition of Fin and aligns with its “Claudeforce” partnership with Anthropic.

For platform teams, this deal points toward a deep overhaul of Salesforce’s outbound communication fabric. Braze operates high-throughput, low-latency ingestion and messaging pipelines designed for real-time mobile and web streaming. Integrating those pipelines with Salesforce Data Cloud and Anthropic-driven agents means automated sales and support workflows will gain native triggers into multi-channel campaigns without custom webhooks and integration middleware.

AI-discovered remote code execution flaws force emergency PHP patches

Source: apnews.com ↗

Critical vulnerabilities discovered with the assistance of AI security tools in PHP prompted emergency downtime across production web fleets this week. Security agencies warned that the flaws could lead to unauthenticated remote code execution (RCE). Several organizations took public properties offline to execute patch cycles before proof-of-concept exploits spread.

The incidents highlight the shift in vulnerability discovery economics. Automated scanners running targeted large models are identifying complex, context-dependent memory corruption and edge-case execution paths in legacy C codebases faster than maintainers can quietly triage them. If you operate PHP runtimes in production, audit your dependencies, verify your upstream runtime versions, and ensure your ingress web application firewalls are blocking malformed input until runtime patches land.

Blacksmith hits $1B valuation as AI agents hammer CI pipelines

Source: techcrunch.com ↗

CI infrastructure startup Blacksmith raised $66 million in a Series C round led by Sequoia, bringing its valuation to $1 billion just two months after its Series B. The platform has logged 17x year-over-year ARR growth and executed more than 90 million CI jobs for engineering teams like Mercury and Granola.

The catalyst is automated software generation. Human engineers commit code periodically; coding agents running in loops commit and test continuously, creating exponential demand on continuous integration runtimes. Blacksmith runs bare-metal compute optimized for quick spin-up times, providing an alternative to stock GitHub Actions runners that bottleneck under agent-driven pull request volumes.

Rethinking developer tooling for coding agents

Source: ines.io ↗

Explosion, the team behind spaCy, is tackling the interface problem between automated agents and developer platforms. As outlined in a detailed breakdown on agent developer UX, Explosion is building Ellf, an agentic NLP development platform that exposes unified interfaces for both human web users and terminal-based coding agents like Claude Code.

The core argument: building SaaS platforms solely around visual dashboards turns developers into “reverse centaurs” - manual cut-and-paste glue between browser UIs and terminal environments. Ellf generates synchronized visual UIs and programmatic CLIs directly from Python functions and type signatures. As autonomous agents become the primary operators of dev tools, platforms must expose typed, deterministic command-line and API interfaces alongside their graphical counterparts.

The enterprise recovery wall: Defining what "good" looks like

Source: cio.com ↗

Modern disaster recovery plans are breaking against configuration drift and automated deployments. An analysis of the CIO recovery problem notes that while spinning up compute across hybrid cloud and SaaS environments is largely solved by Infrastructure as Code, verifying the integrity of the restored system state is not.

Continuous deployments, automated IAM policy updates, and dynamic cross-cloud dependencies mean that snapshots restored from backup often contain the very misconfigurations or dormant backdoors that triggered the incident. Uptime restoration speed is becoming secondary to state attestation. Systems engineers need active dependency mapping and continuous configuration drift baselines before an incident occurs, rather than relying on point-in-time recovery runbooks.

Amazon ECS Gateway simplifies container ingress

Source: aws.amazon.com ↗

AWS launched Amazon ECS Gateway, a managed ingress service that automates the lifecycle of Application Load Balancers for containerized workloads. Available across all commercial regions at no additional charge beyond the underlying AWS resources, the feature provisions ALBs directly from ECS service definitions.

ECS Gateway requires the awsvpc network mode, handles host-based routing natively, and allows up to 25 separate services to share a single ALB with unique DNS hostnames. This eliminates the need for custom ingress controllers, complex Terraform modules, or dedicated load balancers for every internal microservice, directly trimming ALB baseline hourly costs across ECS clusters.

Databricks introduces Agent Memory to slash token overhead

Source: databricks.com ↗

Databricks rolled out Agent Memory in beta within its Agent Bricks ecosystem. Running on Databricks Lakebase - its serverless Postgres layer - and utilizing a custom MemCurrent architecture, the engine maintains long-term user preferences, system state, and context across distributed agent runs.

The system incorporates context-loading plugins, runtime retrieval tools, and an offline “dreaming” pipeline that consolidates disparate session events during idle windows. Databricks claims the tiered architecture tripled agent task completion rates while cutting prompt token consumption by 31% by eliminating the need to replay massive conversation contexts into every LLM request.

Phishing campaigns pivot to GCP "LLMjacking"

Source: wiz.io ↗

Cloud security researchers at Wiz CIRT detailed an emerging attack vector dubbed LLMjacking targeting Google Cloud accounts. Attackers phish corporate credentials and MFA tokens, gain administrative access to cloud environments, and modify resource policies to link external, attacker-controlled GCP projects directly to the victim’s internal Cloud Billing account.

The primary target is quota for high-end LLM inference, specifically Gemini API endpoints. Because the actual compute workloads and prompts execute within the attacker’s foreign GCP project, the victim’s internal Cloud Logging and audit trails show zero prompt contents or model requests. Victims only notice the compromise when unexpected billing alerts spike into hundreds of thousands of dollars within days. Security teams must restrict billing.resourceAssociations.create permissions and enforce org-level policies blocking billing exports to external project IDs.

Gemini enterprise search lands in air-gapped perimeters

Source: cloud.google.com ↗

Google Cloud launched a preview of Gemini-powered enterprise search running fully disconnected on Google Distributed Cloud (GDC) appliances. The on-premises stack runs on Nvidia Blackwell Ultra accelerators, providing complete retrieval-augmented generation (RAG) capabilities inside air-gapped defense and sovereign enterprise perimeters.

The appliance packages document parsers, OCR, audio/video transcription, and a managed vector database into a turnkey private appliance. For systems teams operating under strict sovereignty mandates, this eliminates the need to build, patch, and operate bespoke offline RAG pipelines across detached internal GPU clusters.

High-performance inference: Kubernetes llm-d 1.0 and DeepSeek DeepInfer

Source: cloud.google.com ↗

Infrastructure options for open-weights model inference are solidifying around specialized routing and kernel stacks. Google, Red Hat, and IBM launched llm-d 1.0, a CNCF Sandbox project designed for Kubernetes-native distributed inference. Working alongside backends like vLLM and SGLang on GPUs and TPUs, llm-d adds cache-aware routing, tiered KV caching, and priority flow control to stop long-context agent queries from causing head-of-line blocking on shared serving clusters.

Concurrently, DeepSeek open-sourced its production engine, DeepInfer, under the Apache 2.0 license. Geared toward offline batch inference, DeepInfer launches with support for DeepSeek-V4.1-Flash and requires Linux with GCC 13.0+ and Python 3.12. It targets Nvidia SM100 architectures via CUDA 13 and Huawei Ascend 950 NPUs via CANN 9.2, using TileLang kernels and speculative draft token verification. Production operators should note that DeepInfer omits online components: prefill/decode disaggregation, external KV caching layers, and ingress load balancing must be supplied by the operator.