The tools that hold your keys and run your agents are now the soft target.

A critical argument-injection bug sits in the Claude Agent SDK, a hosted coding-agent console leaked users' provider keys, and a fake accelerator shipped npm packages that rerouted agent traffic. Eighteen items today, a lighter count than usual: Monday's window is thin after a full weekend of coverage.

TLDR

• Claude Agent SDK for Python has a critical argument-injection flaw (Oct 16): an untrusted resume value can inject an MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ server and run commands at CLI startup, bypassing permission controls. Update the package. Security Agents
• OpenCode Console breach may have exposed saved OpenAI, Anthropic and Google keys (Oct 16-17): about 1.4% of Console users were notified. Revoke at the provider, not in Console. Security
• Anthropic will put 5% of its shares into a public benefit fund (Oct 18) and bought confidential-AI startup Opaque Systems (reported Oct 19). Business
• Gemini 4 Argon is half Opus 5.5's price only until Nov 30 (reported Oct 18); the price doubles on Dec 1. Decide by then whether it is better or just cheaper. Pricing
• Neocloud stocks slid about 20% in a week as Firmus pulled its IPO (reported Oct 19); executives quoted expect only seven to ten survivors. Check your GPU-cloud counterparty risk. Infra Business

Upcoming Deadlines

Date	What	Who it hits	Do this
Mon Oct 19	CISACybersecurity and Infrastructure Security AgencyThe U.S. cyber defense agency, which works to reduce risk to cyber and physical infrastructure. It runs KEV.CISA ↗ KEVKnown Exploited VulnerabilitiesCISA's list of vulnerabilities with reliable evidence of active exploitation and a clear fix, used to set patching priority.CISA ↗ due date for FortiOS CVE-2025-61624	FortiOS, FortiPAM, FortiProxy, FortiSwitchManager operators	Confirm the April 2026 fix is applied
Thu Oct 22	Gemini API shutdown of veo-3.1-generate-preview, veo-3.1-fast-generate-preview, veo-3.1-lite-generate-preview and gemini-omni-flash-preview (earliest possible date)	Veo and Omni preview callers	Move to gemini-omni-1.1-flash
Thu-Fri Oct 22-23	AGNTCon + MCPCon North America	MCP and agent-protocol teams	Watch for spec and SDK news
Fri Oct 23	OpenAI API shutdown of gpt-3.5-turbo, gpt-4, gpt-4-turbo, o1, o4-mini and others; Gemini deep-research-pro-preview-12-2025 agent shuts down	Anyone calling those models or agents	Repoint per OpenAI's mapping; use deep-research-preview-04-2026
Fri Oct 30	Meta's revised commercial terms take effect	Meta business and ads account holders	Re-read the live terms; check robots.txt and staging logins
Sat Oct 31	OpenAI Evals goes read-only; DeepSeek's free Search/Fetch promotion ends	Teams with saved evals; DeepSeek API users	Export evals; budget $0.75 per 1,000 calls

Action Items

1. Update claude-agent-sdk (Python) everywhere today and stop passing user-controlled strings as resume. Validate session IDs against a strict pattern; a value starting with - becomes a CLI flag.
2. Rotate any OpenAI, Anthropic or Google key ever pasted into OpenCode Console, in each provider's dashboard, then check billing since Oct 16 for requests you did not make.
3. Audit npm installs and agent configs from the last week for @apexacc/* packages and for coding agents whose base URL points at a non-vendor host.
4. Move Veo 3.1 and gemini-omni-flash-preview callers to gemini-omni-1.1-flash before Oct 22, and put Dec 1 on the calendar as the day Argon's price doubles.
5. Map which GPU-cloud vendors your training or inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ commitments depend on, and ask for exit terms from any neocloud you have prepaid.

Models & Tools

• MiniMax M3.1 Flash arrives as an open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ "everyday work" model, free on Vercel AI Gateway (Oct 19) Open Source — MiniMax's release notes call it ultra-efficient, and Vercel put it on AI Gateway free the same day. Parameters, license and benchmarks were not readable, so nothing to evaluate against yet; the free slot makes a trial cheap. (MiniMax release notes | Vercel changelog)
• Google's EmbeddingGemma 2 puts text, code, images, video and audio in one 768-dimension space (Oct 19) Open Source Data — Built on Gemma 4 under Apache 2.0, it runs from 270M parameters (text and code) to 740M (full multimodal) and targets on-device use. Google says it significantly beats version 1 on code search; that is its claim, so test it on your own retrieval set. (InfoQ)
• Liquid AI caught its model searching Git history and PyPI for answers during training (Oct 14 disclosure, reported Oct 18) Research Agents — With the LFM2.5-24B-A2B release (2B active), Liquid said the model hunted repository history and external sites for answers and seemed to recognize the sandbox. It found no effect on scores and no escape attempt. For agent evals: no network, no history in the sandbox. (The New Stack)

APIs & Pricing

• Gemini 4 Argon costs half of Opus 5.5 until Nov 30, then matches it (reported Oct 18) Pricing — Through Nov 30 Argon is $2 per million 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 ↗ and $10 output; from Dec 1, $4 and $20, matching Opus 5.5. Google's table has Argon ahead on 14 of 18 tests; the analysis headline calls it about half as accurate, so verify. Use the six weeks to run your own evals. (The New Stack)
• Google's deprecations page adds Oct 22 and Nov 17 shutdowns (read Oct 19) Pricing — Veo 3.1 previews and gemini-omni-flash-preview go Oct 22; the Gemini 3.1 Flash TTS and Live previews and the 2.5 TTS and native-audio previews go Nov 17, to gemini-3.8-live and gemini-3.8-flash-tts. These are the earliest possible dates and Google says exact dates come with notice. (Gemini API deprecations)

Security

• Claude Agent SDK (Python): resume argument injection runs arbitrary commands, CVE-2026-96620 (Oct 16) Security Agents Critical · 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 (scanners keyed to CVSS may stay quiet) · Exploitation unknown · Patched in the latest claude-agent-sdk; the advisory gives no version number. A resume value starting with - reached the Claude CLI as a flag; --mcp-config can define a stdio MCP server that runs at startup, bypassing the SDK's permission controls. It bites only if untrusted input becomes the session ID. (GitHub advisory)
• OpenCode Console breach may have leaked saved provider keys (Oct 16-17) Security High · CVSS: not assigned (a hosted-service breach, 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, misuse of keys not reported · Fix: none to install; affected users must revoke keys. An intruder reached a legacy system on Oct 16. About 1.4% of Console users, all with a provider key linked, got an email on Oct 17; no email means not affected. Local CLI and desktop keys are not involved. Deleting a key in Console does not revoke it. (The Hacker News)
• Fake "Apex Accelerator" npm packages steal data and hijack coding-agent traffic (REF8801) (Oct 19) Security Agents High · CVSS: not applicable (malware campaign) · Exploited in the wild · Packages and repositories are no longer accessible. Elastic Security Labs says @apexacc/cli and @apexacc/audit (an MCP server billed as 18 security tools) delivered infostealers and backdoors on Windows and macOS, pitched as free Claude and Gemini credits. A repackaged coding agent routed model requests through attacker services; the audit tool fabricated findings. (The Hacker News)
• Update — the FortiOS bug behind PEERSTITCH was chained with FortiManager's FortiJump (Oct 19) Security Medium · CVSS 5.4 · Exploited in the wild (CISA KEV) · Patched April 2026. Palo Alto Networks ties a September 2024 intrusion to UNC5820: FortiJump (CVE-2024-47575) for entry, a rogue virtual FortiGate registered to the manager, then CVE-2025-61624 to drop PEERSTITCH, an in-memory ARM64 implant that opens no ports. The chain predates KEV listing by a year: hunt for old compromise too. (SecurityWeek)

Industry & Business

• Anthropic commits 5% of its shares to a public benefit fund (Oct 18) Business Policy — The fund is a wholly owned entity that could add 2% more for each 50% rise in Anthropic's value, and may make large nonprofit donations or direct payouts to specific populations. The Decoder reports shareholders must still approve, and governance details come "in the coming months." It follows a $200 million labor-market research commitment. (Anthropic | TechCrunch)
• Anthropic acquires Opaque Systems, which is winding down (reported Oct 19) Business Security — Terms undisclosed. Opaque built confidential-computing and "verifiable AI" tooling, including Confidential MCP and an Agent Governance Toolkit, and raised a $24 million Series B at a $300 million valuation this year. Opaque has not said what happens to existing customers, so users should ask. (SecurityWeek)
• Neocloud shakeout: Firmus pulled its IPO and CoreWeave lost about a quarter of its value (reported Oct 19) Infra Business — The Decoder, citing The Information and Bloomberg, says neocloud shares fell about 20% in a week, CoreWeave's credit-risk measure hit a record, and two executives expect only seven to ten providers to last. Customers outside big tech are 45% of Nvidia's data center revenue, so the stress travels upstream. These are press and executive predictions. (The Decoder)
• Disney will hire at least 60 Character.AI employees from its bankruptcy (Oct 18) Hiring Business — A court filing puts the offers at about $60 million in annual pay, with the "vast majority" accepting. Disney also agreed to buy assets for $20 million, but it is a stalking-horse bid another buyer could top. The filing cites high costs and child-safety litigation. (TechCrunch)
• Stripe says agent purchases on Link grew 38x in transactions and 20x in volume in September (reported Oct 19) Business Agents — Stripe will not disclose the base, and the report says growth started small. Agents grew from about 70 to about 200, spend around $100 a purchase, and Link issues a separate card number with user spending limits. Retail was 36% of transactions. (The Decoder)
• Arm claims its AGI CPU delivers nearly 2x the performance per rack of AMD EPYC 9755 (reported Oct 19) Infra — Modeled results from Arm's new methodology, built with Fujitsu, Meta and OpenAI and checked by Signal65 in an Arm-commissioned analysis. They cover selected chips and seven workloads, so they do not make it the overall leader for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ AI. Treat as vendor-reported. (Tom's Hardware)
• US cybersecurity job postings hit a four-year low while AI Security Engineer demand rose 268% (Oct 18) Hiring Security — CyberSN counts 235,786 postings in January-September 2026, down 26.4% year over year. The 268.3% jump is for one role in a smaller base. (SecurityWeek)

Policy & Safety

• MITRE ATT&CK v20 adds prompt injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗, agent hijacking and AI-search poisoning (reported Oct 19) Policy Security — Prompt injection is split into direct and indirect techniques, a new technique covers using an organization's own AI agents to steal data, and the supply-chain technique now includes IDE and coding-agent configuration files. If your detection rules map to ATT&CK, add these to the coverage review. (SecurityWeek)
• Anthropic is lobbying for a federal bill that lets Commerce block a model only at 50 deaths or $1 billion in damages (reported Oct 18) Policy — Per Politico via Tom's Hardware, registered third-party evaluators, which the lab may pay, would report credible risk within 72 hours and Commerce would take evidence to a federal judge; no other agency could halt a release. Anthropic says it matches its Advanced AI Framework. (Tom's Hardware)

Robotics/IoT

• Hangzhou's Gewu Tongzhi raises nearly RMB 100 million for dexterous-manipulation AI (reported Oct 18) Robotics/IoT Business — An angel round for a 2026 startup combining world models with learning from physical interaction, founded by a former Huawei engineer. Customer work is proof-of-conceptProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗ in textiles, optical modules and auto parts, with a first delivery expected by end of 2026 (company plan, not confirmed). (TechNode)