JadidJust-in AI Developments, Incidents & Deadlines
May 2026
SMTWTFS12345678910111213141516171819202122232425262728293031

AI news, Sunday, May 10, 2026

11 stories

TLDR

  • DALL-E 2 & 3 shut down in 2 days (May 12) — Final window to migrate image-generation pipelines to gpt-image-1; API shape changes make this a non-trivial swap.
  • Spring AI CVE-2026-41705 (May 9) — Filter-expression injection in MilvusVectorStore; CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 8.6; upgrade Spring AI to 1.0.7 / 1.1.6 now.
  • xAI Grok 4.3 live on API (May 5) — #1 in agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ tool calling; 1M tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ context at $1.25/$2.50 per MTok; 8 legacy Grok models retire May 15.
  • IBM Think 2026: watsonx Orchestrate becomes an agentic control plane (May 5) — Deploys agents from IBM, Anthropic, OpenAI, or custom sources with consistent policy enforcement; IBM Confluent real-time data layer also ships.
  • Meta 8,000 layoffs execute May 20 — Announced April 23; 10 days out, teams across Reality Labs, Facebook, recruiting, and ops affected; 6,000 open reqs also canceled.

Action Items

Things your engineering team should look at or do this week.

  • DALL-E 2 & 3 shut off Monday, May 12 — last 48 hours — If you haven't tested gpt-image-1 yet, do it today. The request/response shape differs: resolution options changed (1024×1536 replaces 1792×1024), billing moved from per-image to per-token (~1,000 output tokens per 1024×1024 medium quality image). A naive model-name swap will break clients. (DEV.to migration guide | OpenAI deprecations)
  • Upgrade Spring AI for CVE-2026-41705 — Unsanitized document IDs in MilvusVectorStore#doDelete() allow arbitrary filter-expression injection into Milvus queries. No public PoCProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗ yet, but the fix is clean: Spring AI 1.0.x → 1.0.7, Spring AI 1.1.x → 1.1.6 (released May 8). (Spring Security Advisory | TheHackerWire)
  • Migrate off retired xAI Grok models before May 15 noon PT — grok-4, grok-4-fast, grok-4-1-fast, grok-code-fast-1, grok-imagine-image-pro, and three other variants will return errors after May 15. Replacement is grok-4.3, which brings 1M token context, improved tool calling, and lower pricing ($1.25/$2.50 per MTok). (xAI retirement guide)

Models & Tools

  • Grok 4.3 now on xAI API (May 5) [Agents] — xAI's fastest model to date tops the Artificial Analysis agentic tool-calling and instruction-following leaderboards, and ranks #1 in enterprise domains (case law, corporate finance) on Vals.AI. Supports 1M token context, function calling, web search, code execution, and remote MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server connections. Priced at $1.25/$2.50 per MTok — competitive at that context size. Eight legacy models retire May 15. (xAI announcement | Artificial Analysis)

  • IBM watsonx Orchestrate next gen: multi-agent control plane (May 5) [Agents] [Infra] — Announced at Think 2026, the next generation of watsonx Orchestrate (private preview) evolves from a standalone automation tool into an agentic control plane that can deploy, govern, and monitor agents from any source — IBM, Anthropic, OpenAI, or custom builds — with consistent policy enforcement. Paired with IBM Confluent (real-time Kafka/Flink data streaming) and IBM Sovereign Core (policy-at-runtime governance for regulated environments). For teams evaluating enterprise agent orchestration that needs auditability across vendors, this is the most compliance-focused option to evaluate. (IBM Newsroom | IBM announcements)

  • Gemma 4 Multi-Token Prediction drafters ship (recent) [Open Source] [Infra] — Google released MTP drafters for the Gemma 4 family delivering up to 3× inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ speedup with no degradation in output quality or reasoning. Gemma 4 (Apache 2.0, four model sizes from 2B to 31B, 256K context, vision + audio) has already surpassed 60M downloads since its April launch. The MTP drafters make self-hosted Gemma 4 a serious option for latency-sensitive inference workloads. (Google Blog – MTP Drafters | Gemma 4 overview)

APIs & Pricing

  • DALL-E 2 & 3 retirement: May 12 (update) [Trending] — Zero days remain after Monday. After May 12, all calls to dall-e-2 and dall-e-3 model IDs will fail. Replacement: gpt-image-1 and gpt-image-1-mini. See Action Item above for migration details. (OpenAI deprecations page)

  • xAI: 8 Grok models retire May 15 [Agents] — Affected: grok-4, grok-4-fast, grok-4-1-fast, grok-4-0709, grok-3, grok-code-fast-1, grok-imagine-image-pro, and related aliases. Replace with grok-4.3. Migration deadline is nine days from the May 6 announcement — an unusually tight window. (xAI docs | Migration guide)

Security

  • Spring AI CVE-2026-41705: filter-expression injection in MilvusVectorStore (May 9) [Data] — MilvusVectorStore#doDelete(List<String>) passes unsanitized document IDs directly into a Milvus filter expression, allowing injection of arbitrary query logic. CVSS 8.6 (High). No public PoC at time of disclosure but the injection pattern is straightforward for anyone with write access to document IDs at ingestion time. Affected: Spring AI 1.0.x (all versions) and 1.1.x (all versions). Fix available: Spring AI 1.0.7 and 1.1.6 (released May 8, one day before CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ publication). Note: 5 Spring AI CVEs were disclosed April 27 — if you haven't already audited your Spring AI upgrade path, the vector store layer is a recurring injection surface. (Spring Security Advisory | TheHackerWire | HeroDevs Spring AI CVE roundup)

Industry & Business

  • Meta 8,000 layoffs execute May 20; 6,000 open reqs canceled (announced Apr 23) [Hiring] [Trending] — With execution 10 days out, this is the most significant single-round tech workforce reduction of 2026. The cuts are structural rather than performance-based: Meta is reorganizing into AI-focused "pods" and moving engineers across Reality Labs, Facebook, recruiting, sales, and global ops into the Applied AI organization. Affected US employees get 16 weeks base pay plus two weeks per year of service, and 18 months of health coverage. Additional cuts are planned for H2 2026. Context: Meta's 2026 capex is $115B+, almost entirely for AI infrastructure. (The Next Web | CNN Business)

  • NVIDIA financing the entire AI supply chain via equity stakes (May 9) [Infra] [Trending] — CNBC reports NVIDIA has now crossed $40B in equity investments across the AI stack — cloud providers, neoclouds, chip designers, and infrastructure companies — an explicit strategy to ensure the AI supply chain runs on NVIDIA hardware and that capacity scales with demand. Meanwhile, Wall Street signals a "changing of the guard": Intel, AMD, and Micron are all outperforming NVIDIA YTD as the inference buildout broadens beyond GPU-only workloads, and Corning is seeing historic demand for fiber-optic cables in new data centers. (CNBC – NVIDIA equity strategy | CNBC – chip landscape shift)

AI in Production

  • Enterprise GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ utilization: 5% average, $401B problem (ongoing) [Infra] — VentureBeat surfaces a stark figure: average enterprise GPU utilization is 5%, representing an estimated $401B in stranded compute. The root causes are poor scheduling, idle reservation, and mismatched inference vs. training workloads. Inference now accounts for ~two-thirds of all AI compute in 2026, but most enterprise procurement was designed for training. The practical implication: before adding capacity, audit your utilization metrics — most teams have untapped headroom. (VentureBeat)

Policy & Safety

  • 2026: The Year of AI-Assisted Attacks — Hacker News threat brief (May 2026) [Security] [Trending] — The Hacker News published an industry-wide threat assessment: AI-assisted attacks are no longer theoretical. The piece catalogs documented 2026 incidents including large-scale indirect prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ attacks in the wild (Unit 42, March 2026), AI-generated phishing campaigns operating at scale, and the OWASP ranking of prompt injection as LLM01:2025 — the top vulnerability class. Key stat: researchers recorded over 461,000 prompt injection submissions in a single public dataset, with success rates of 50–84%. (The Hacker News)

Paper Spotlight

  • "The Art of Building Verifiers for Computer Use Agents" (May 2026) — Microsoft Research and Browserbase published research on how to reliably score whether a computer-use agent actually succeeded at a task (not just whether it reported success). They built a Universal Verifier based on four principles that achieves human-level agreement rates with near-zero false positives. This is a direct solution to the agent reliability problem: agents that can't be verified can't be trusted in production. Practically useful for any team building computer-use or web-automation agents. (Microsoft Research / Browserbase)

Type to search · ↑↓ to move · enter to open