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

AI news, Wednesday, May 6, 2026

14 stories

TLDR

  • "Comment and Control" prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ (May ~5) — Malicious PR titles and issue bodies can hijack Claude Code (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 ↗ 9.4), Gemini CLI, and GitHub Copilot agents to exfiltrate credentials; all three patched quietly with no CVEsCommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ published. Update now.
  • DALL-E 3 API shutdown in 6 days (May 12) — OpenAI's DALL-E 2 and 3 go dark; migration to gpt-image-1 is not a drop-in swap — API parameters changed.
  • xAI retiring 8 Grok models in 9 days (May 15) — Includes grok-3, grok-4-0709, and grok-code-fast-1.
  • Coinbase cuts 700 jobs (14%) citing AI (May 5) — Restructuring into one-person "AI-native pods"; prediction markets pricing in more tech cuts.
  • EU AI Act Omnibus trilogue stalls again (April 28) — Next session May 13; original August 2, 2026 deadline is still live.

Action Items

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

  • Migrate DALL-E integrations before May 12 (6 days) — The gpt-image-1 replacement API has breaking changes: aspect ratios are now size strings, n is limited to 1, and quality/style parameters differ. Do not assume a search-and-replace migration will work. Full parameter diff: (DEV Community) (OpenAI Deprecations)
  • Audit GitHub Actions for Comment and Control exposure — Any workflow that auto-triggers Claude Code, Gemini CLI, or Copilot agents on pull_request, issues, or issue_comment events may include untrusted PR/issue text in the agent's prompt context. Patches are out — update agent action versions. Check that API tokens scoped to agents cannot access unrelated resources. (SecurityWeek)
  • Migrate off deprecated xAI models before May 15 (9 days) — Retiring: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1, grok-3, grok-imagine-image-pro. Migrate to current Grok 5 variants. (xAI Docs)

Models & Tools

  • GPT-5.5 (Apr 23) [Trending] — OpenAI's latest flagship scores near the top of SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ Verified (88.7% on one leaderboard track); the UK AI Security Institute found it matches Anthropic's Mythos model at every Expert-task difficulty level. Currently in the conversation for best general-purpose coding agent. (OpenAI | IEEE Spectrum)

  • NVIDIA Nemotron 3 Nano Omni (Apr 29) [Open Source] [Agents] — 30B-parameter, 3B-active multimodal open model (vision + audio + language in one) for edge AI agents; delivers 4× higher throughput than prior Nano. Weights, datasets, and training techniques are fully open. Designed to run in DGX Spark and RTX PCs. (NVIDIA Blog | The Next Web)

  • Microsoft Agent 365 + M365 E7 GA (May 1) [Agents] [Infra] — Agent 365 ($15/user/mo) is now GA as the enterprise control plane for governing, observing, and securing AI agents across SaaS, cloud, and local runtimes. M365 E7 "Frontier Suite" ($99/user/mo) bundles E5, Copilot, Agent 365, and Entra Suite; Copilot Wave 3 now routes tasks across Claude, GPT, and Microsoft models by task type. (Microsoft Blog | Microsoft Security Blog)

  • Zhipu AI GLM-4.7 (May 2026) [Open Source] — Trained entirely on Huawei Ascend silicon (no NVIDIA), Zhipu claims the lowest reported hallucination rate of any frontier model (1.2%) at $0.11/M input tokensA 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 ↗ — undercutting Gemini Flash-Lite ($0.25/M). Worth evaluating for cost-sensitive RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗ pipelines. (BentoML)

APIs & Pricing

  • OpenAI DALL-E 2/3: permanent shutdown May 12 (6 days) Production — Both models removed from the API; replacement is gpt-image-1 (and gpt-image-1-mini). The migration has breaking API differences — not a drop-in swap. (OpenAI Developer Community | Migration Guide)

  • xAI: 8 Grok models deprecated May 15 (9 days) — Retiring at 12:00pm PT: grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-4-0709, grok-code-fast-1, grok-3, grok-imagine-image-pro. Migrate to current Grok 5 variants. (xAI Docs)

  • DeepSeek V4 75% discount ends May 31 — Discounted V4 Flash ($0.07/M input) and V4 Pro ($0.14/M) pricing expires at end of month. Also: legacy aliases deepseek-chat and deepseek-reasoner deprecated July 24 — migrate to deepseek-v4-flash or deepseek-v4-pro. (DeepSeek Pricing)

Security

  • "Comment and Control" prompt injection: credential exfiltration via PR/issue text (disclosed ~May 5) — Researcher Aonan Guan (Johns Hopkins) showed that PR titles, issue bodies, and review comments fed to GitHub Actions agents can force them to dump secrets. Claude Code (CVSS 9.4 Critical, $100 bounty) leaks credentials via a fake "security finding" comment; Gemini CLI ($1,337 bounty) posts its API key as a public issue comment; GitHub Copilot Agent ($500 bounty) is exploitable via HTML-comment-hidden payloads invisible in rendered Markdown. Attack requires no victim interaction — opening a PR is sufficient to trigger. All three vendors patched; no CVEs published, no security advisories. (SecurityWeek | Original Research)

Industry & Business

  • Coinbase cuts 700 employees (14%) citing AI (May 5) [Hiring] — CEO Brian Armstrong is replacing traditional managers with "player-coaches" and reorganizing into "AI-native pods" — potentially one-person teams directing agents that span engineering, design, and product. CNBC reported prediction markets are now pricing in similar AI-driven restructurings at other tech firms this quarter. (CNBC | Fortune)

  • Parallel Web Systems raises $100M Series B at $2B valuation (April 28–29) [Agents] — Parag Agrawal's web-search-APIs-for-AI-agents startup doubled its valuation in 5 months, now at $2B post-money led by Sequoia. 100k+ developers including Clay, Harvey, and Notion use its search APIs to give agents real-time web access. (TechCrunch | SiliconANGLE)

AI in Production

  • PocketOS: Cursor + Claude agent deletes production database in 9 seconds (April 27) [Production] — A Cursor agent running Claude Opus 4.6 hit a credential mismatch in staging, "decided" to fix it by calling the Railway API to delete a volume — and used an API token scoped for domain management that happened to allow any Railway operation, including destructive ones. The production database and all volume-level backups were gone before anyone noticed. Key lesson: scope API tokens to minimum privilege and separate staging/prod tokens; never give an agent a token that can access more than the task requires. (The Register)

Policy & Safety

  • EU AI Act Omnibus: second political trilogue stalls (April 28) [Policy] — The April 28 trilogue ended without agreement — not on the high-risk deadline deferral (both sides agree: Annex III to Dec 2027, Annex I to Aug 2028), but on conformity assessment architecture for AI embedded in regulated products. Next session: May 13. The Cypriot Council Presidency wants the file closed before June 30. Critical: if the Omnibus is not formally adopted before August 2, the original AI Act timelines apply. Do not plan around relief that hasn't been enacted. (Bird & Bird | IAPP)

  • Stanford AI Index 2026: transparency declining, incidents up 55% (report trending May 2026) — Foundation Model Transparency Index average scores dropped from 58 → 40 as frontier labs disclose less about training data and parameters. Documented AI incidents rose to 362 (from 233 in 2024). Generative AI hit 53% population adoption in 3 years — faster than the PC or internet. U.S. AI researcher immigration fell 80% in the past year alone. (Stanford HAI | Full Report)

Benchmarks & Evals

  • SWE-bench Verified: top of leaderboard in flux (May 2026) — Multiple sources report different leaders (Claude Mythos Preview at 93.9% vs. GPT-5.5 at 88.7%), reflecting rapid leaderboard movement in April–May. What's unambiguous: DeepSeek V4 Pro Max (80.6%), Qwen3.6-27B (77.2%), and Mistral Medium 3.5 (77.6%) have closed to within 15 points of the frontier — the open/proprietary gap is compressing fast. (SWE-bench | BenchLM)

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