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

Disclosure is outrunning patching, and Anthropic's answer is to prove code correct instead of testing it harder.

20 stories · transcript

PHP flagged a third remote-code path with no fix scheduled, Pillow 13.0.0 closed a batch of critical image-parsing flaws, and Next.js has an unauthenticated RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ with no workaround. Claude produced machine-checked proofs for 123 open-source libraries, with the limits Anthropic states.

TLDR

  • PHP warns of a third RCE, CVE-2026-115560, two days after the PHP2Shell emergency release (Oct 23), in a narrow session configuration, with no fix until the next routine release. Set one ini value now. Security
  • Next.js GHSA-673q-mjm6-r5xr is an unauthenticated RCE in use cache functions that return typed arrays (Oct 22), with no workaround. Security
  • Pillow 13.0.0 fixes a batch of critical-rated flaws, including an EPS filename command injection, CVE-2026-112900 (Oct 16). Security
  • Anthropic and Theorem use Claude to formally verify 123 OSS-Fuzz libraries (Oct 23), with about 180 million lines of proof code and strict stated limits. Research Agents
  • OpenAI's Decisions API is generally available with gpt-6-luna (Oct 22), answering TypeSafe's Jev at roughly 100 to 180 ms per call. Production Pricing

Upcoming Deadlines

Date What Who it hits Do this Source
Thu-Fri Oct 22-23 AGNTCon + MCPCon North America MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ and agent-protocol teams Watch for spec and SDK news Linux Foundation
Fri Oct 23 OpenAI API shutdown of o1, o1-pro, o3-mini, o4-mini, gpt-4, gpt-4-turbo, gpt-4.1-nano, gpt-3.5-turbo, gpt-image-1 and matching fine-tunes Anyone calling those models by API Repoint per OpenAI's mapping OpenAI
Fri Oct 23 Gemini API deep-research-pro-preview-12-2025 agent shuts down Deep Research agent users Change agent in interactions.create Gemini API
Fri Oct 23 MiniMax M3.1 Flash open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ expected (SGLang added day-0 support Oct 22; license not stated) Self-hosters of open models Line up an eval SGLang
Sat Oct 31 OpenAI Evals goes read-only; DeepSeek's free Search/Fetch promotion ends 4:00 PM UTC Teams with saved evals; DeepSeek API users Export evals; budget $0.75 per 1,000 calls OpenAI · DeepSeek
Sat Oct 31 Mistral retires zai-glm-5-2 Callers of that model Move to zai-glm-5-3 at the same price Mistral

Action Items

  1. Upgrade Pillow to 13.0.0 or later today and rebuild every image, wheel and base layer that bundles an older copy, then remove the gs binary from hosts that process untrusted uploads. Python 3.10 installs must move to 3.11 first.
  2. Set session.upload_progress.enabled=0 on every PHP host this week, and confirm the PHP2Shell fixes are live on 8.5.12, 8.4.27, 8.3.36 or 8.2.35.
  3. Grep your Next.js App Router code for use cache functions that return typed arrays, and upgrade to the patched release named in GHSA-673q-mjm6-r5xr today, since the advisory says to upgrade even without such a function.
  4. List every agent that signs x402 payments and confirm it runs the August 18 SDK release or later with a per-payment spending cap, because npm audit and PyPI show no advisory for the older versions.
  5. Export saved OpenAI Evals data before Oct 31, ahead of the read-only switch and the Nov 30 shutdown.

Models & Tools

  • OpenAI makes the Decisions API generally available with gpt-6-luna (Oct 22) Production Pricing The API returns a choice plus a confidence score and now accepts up to 400 questions per request, double the beta limit. OpenAI says latency fell from 1 to 3 seconds to 100 to 180 ms; Speak reports 1.4 seconds down to 0.13. It is a reconfigured Luna, not a purpose-trained decision model. (TechCrunch | OpenAI)
  • Anthropic and Theorem use Claude to formally verify 123 open-source libraries (Oct 23) Research Agents Claude turned OSS-Fuzz test programs into proofs about the compiled machine code, about 180 million lines checked by Rocq. Proofs hold only for one processor architecture, single-threaded code and inputs under 64 MB, and single proofs run to hundreds of millions of 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 ↗. Vendor-reported. (The Decoder | Anthropic)
  • Google previews Agent Compute Engine, a hardware-isolated microVM for every agent (Oct 22) Infra Agents Preview arrives later this year. Idle VMs pause and resume; Google says 200,000 agents resumed in under 500 ms and one server holds ten times as many agents. E2B, Daytona and AWS AgentCore already sell similar sandboxes. (The New Stack | The Decoder)
  • DeepSeek V4.1-Flash-1017 undercuts Claude Haiku 5.5 in a hands-on test (Oct 17) Open Source Pricing The 552B-parameter open-weight MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ activates 16B and claims 42.3 on Terminal-Bench 4.0 against Haiku's reported 39.2; both are vendor-reported. DeepSeek's off-peak price is $0.15 in and $0.60 out per million tokens, double at peak. Tom's Hardware finds it more hallucination-prone than proprietary rivals. (The New Stack | Tom's Hardware)

APIs & Pricing

  • OpenAI adds 12-hour prompt caching, Ultrafast mode for GPT-6.1 Sol, and active-turn sandbox billing (October) Pricing Twelve-hour cache writes cost 2.5 times the input price, so they pay off only when requests reuse a prefix across hours. Ultrafast (service_tier: "ultrafast") cuts time between tokens for all API users. Hosted-sandbox compute no longer bills idle time between turns. (OpenAI)
  • OpenAI donates Open Responses, and Anthropic donates Agent Skills, to the AgenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ AI Foundation (Oct 22) Open Source Agents Open Responses is a provider-neutral API format that Vercel AI Gateway and Databricks already implement. Anthropic says over 70 platforms use Agent Skills. They join MCP, A2AAgent2AgentAn open protocol, started by Google and now under the Linux Foundation, that lets agents from different vendors work together.A2A Protocol ↗ and AGENTS.md under one Linux Foundation roof. (The New Stack | Anthropic)

Security

  • PHP CVE-2026-115560: a third RCE path, with no emergency fix (Oct 23) Security SEVERITY: Unrated · 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 ↗: not yet assigned, so scanners will be silent · Not exploited · Not fixed in 8.5.12, 8.4.27, 8.3.36 or 8.2.35. It needs session.serialize_handler switched to php via ini_set() before session_start() with upload progress on. Mitigation: session.upload_progress.enabled=0. PHP2Shell probing meanwhile reached millions of attempts. (The Hacker News | PHP advisory)
  • Next.js GHSA-673q-mjm6-r5xr: unauthenticated RCE in use cache functions (Oct 22) Security Critical · CVSS: not yet assigned, no CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ · Exploitation unknown · Fixed version not stated in the advisory text. Affects App Router apps with Cache Components where a use cache function returns multiple typed arrays. There is no workaround, and the maintainers say to upgrade even without such a function. (GitHub Advisory)
  • Pillow 13.0.0 fixes CVE-2026-112900, an EPS filename injection into Ghostscript, and a critical batch (Oct 16) Security Critical · CVSS 9.2 (GitHub CVE record; maintainers say moderate) · Not exploited known · Patched in 13.0.0. Needs gs installed and an attacker-controlled filename. Also fixed: a TIFF in-memory save overflow with a public proof of conceptProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗, and a JPEG thumbnail overflow found by Anthropic using Claude. (The Hacker News | GitHub Advisory)
  • x402 SDKs let a malicious API request an agent's full token balance (Oct 22) Security Agents Critical · CVSS: not yet assigned, no advisory published · Not exploited reports · Fixed in the Aug 18 releases. V12 found 25 flaws; the default client had no spending cap. Because the foundation filed no advisory, npm audit and PyPI stay silent for older versions. (The Hacker News)

AI in Production

  • US banks lock out AI agents that log in with customer credentials (Oct 22) Customer AI Agents Morgan Stanley blocks agents entirely; JPMorgan, Bank of America, Wells Fargo and Citi are restricting them, per Axios. Banks prefer read-only access, and most cannot tell an agent acting for a customer from one acting alone. (The Decoder | Tom's Hardware)
  • 1Password for Agents signs an agent in without exposing the password (Oct 22) Agents Production The agent proves its identity, the user approves each login, and 1Password fills the credentials. It supports Claude Duet, ChatGPT Work, Grok Bot, Meta's Muse and Town. Worth studying as the pattern banks are asking for. (TechCrunch)
  • CyberCartel used Claude, Grok and Moonshot's agent against 400-plus Latin American institutions (Oct 23) Security Agents Straiker says the group bypassed refusals by claiming a test environment, then stole passwords and customer data. Anthropic has blocked the access; Straiker is still trying to reach Moonshot. (Tom's Hardware)
  • Harvard study of 718 firms: agents lift output, assistants barely do, and review absorbs the gain (Oct 22) Research Production Agents raised lines of code 30% and pull requests 23% but resolved issues only 12% more; assistants did not raise resolved issues. AI code adds downstream review work. (InfoQ)

Industry & Business

  • OpenAI is reportedly raising up to $30B at a $1.25T pre-money valuation (Oct 22) Business Reported by CNBC, with terms unsettled. The FT says demand was weaker than past rounds and the target was cut from $1.4T. Coatue and UAE investors lead, possibly paired with UAE compute. (TechCrunch | The Decoder)
  • DeepSeek reportedly raises about $15B days after the US federal ban (Oct 22) Business Bloomberg's report names CATL and Tencent as potential investors. It follows a June round of over $7B at a $50B valuation. Unconfirmed by DeepSeek. (TechCrunch | SCMP)
  • Uber puts all 30,000 employees on ChatGPT Work (Oct 22) Customer AI Business A 1,000-person pilot preceded it, and 80% of early users were non-engineers. Uber has not fixed how staff will use it. (TechCrunch | OpenAI)

Policy & Safety

  • Lava finds 32 export-barred Nvidia B300 GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ inside China Telecom's network (Oct 23) Policy Infra The Information reported the finding. Nvidia called diversion "a drop in the bucket" and did not deny it. (Tom's Hardware)
  • A New Mexico court orders Meta to pay $25B over 43 million privacy violations (Oct 22) Policy The judge found Meta falsely told users they controlled their data. The case stems from Cambridge Analytica; an appeal is expected but was not reported. (The Hacker News | SCMP)

Benchmarks & Evals

  • Letta's FrontierMemory finds agent harnesses learn from experience very unevenly (Oct 22) Research Agents Letta Dream scored 73.3%, Prime Agent 60%, OpenAI Agents SDK 40%, Claude Managed Agents 33.3% and Hermes Agent 26.7%. Letta wrote the benchmark and places first on it, so treat the ranking as vendor-reported. (The New Stack)

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