AI Agent Security Cracks | AI Field Notes #103

AnalysisA model whose maker says it beats the previous top tier while charging the budget price is the kind of move that forces every rival to answer. DeepSeek, the Chinese lab whose cheap models rattled US firms in 2025, opened public access to V4.1 Flash around September 10, keeping the same developer endpoint and beta pricing as V4 Flash while adding native image and video understanding. The company says V4.1 Flash beats its own V4 Pro on quality, speed, and cost. For anyone choosing an API to build on, the price of capable inference, the cost of running a model to serve users rather than training it, fell again.

AnalysisA safety fence any passing request can open is not a fence. The security firm VulnCheck published details on September 8 of a flaw in DeepSeek Harness, an open-source tool for running AI coding agents, and scored it 9.4 out of 10 on the standard severity scale. The tool's local interface trusted a client-supplied Host header, so a spoofed value let an unauthenticated attacker set an agent's session to danger-full-access, switching off the sandbox and the approval prompts, then read stored conversations. No API key or model call needed. DeepSeek patched it on August 27, so anyone running an older build is exposed.

AnalysisThe demo everyone feared now has an incident report. Google's Threat Intelligence Group said on September 8 that one financially motivated attacker built an autonomous credential-theft operation in under six hours from off-the-shelf parts: an AI coding chatbot, a prompt, and a set of markdown playbooks that drove the scanning and harvesting on their own. IP rotation and troubleshooting ran with no human at the keyboard. The system later surfaced as a dashboard managing more than 23,800 harvested secrets, including cloud and AI-service keys, drawn from Google's Q3 threat tracker and its Mandiant incident-response arm. The barrier to a professional-grade attack is now a weekend project and a chatbot.

AnalysisWhen the cost of making a browser extension drops to a prompt, the review queue breaks first. Microsoft said on September 8 that submissions to its Edge Add-ons store have surged as AI-assisted coding lets more people assemble extensions fast, straining the pipeline and leaving developers waiting longer for approval. Its fix is more automation: automated validation checks at submission and a 15-day refresh on the Featured badge to surface decent work sooner. The company insists quality standards hold. The part it did not say out loud: automated review now meets automated submission, and low-effort AI extensions arrive faster than any human reviewer reads them.

AnalysisAdobe's answer to the video-model race is to stop making you choose a tab. It unveiled a Generative Media workspace on September 8, timed to the IBC broadcast trade show, that lets editors generate clips, sound effects, music, and ambient audio straight from gaps in the Premiere Pro timeline. It puts five competing models side by side: Adobe's own Firefly plus Google Veo, Runway, Kling, and Luma. Generate Video and Generate Sound Effects ship now; music and soundscape are in beta. The bet is that owning the workflow, where the model is just a dropdown, is what keeps editors inside Premiere.

AnalysisTwo years of copyright suits pushed a company to finally do the thing it refused: pay for the training data. Suno, the AI music startup, released v6 on September 9, its first models built on licensed catalogs from Warner Music Group, BMG, and Believe, with revenue sharing for rightsholders and an opt-in for individual artists who want in. The lineup runs three tiers, including a free v6-mini. It arrives the same season music publishers press a separate suit against Anthropic over unlicensed lyrics, drawing the line of the moment: license the catalog and share the money, or fight it out in court.

AnalysisWith no federal rule on AI in classrooms, a teachers union wrote one into a contract. The American Federation of Teachers (AFT), its New York local, and Microsoft announced a National AI Safety and Privacy Standard on September 9 that school districts can bolt onto their Microsoft agreements as enforceable terms. It bars Microsoft from training AI on student or educator data, from selling that data or using it for ads, and from letting its AI make school decisions without a human in the loop. Districts keep control of what is retained and deleted. The union says it is still pressing Anthropic and OpenAI to sign the same terms.

AnalysisThe contracts that put frontier AI labs inside the Pentagon are public now, and one line stands out. The Intercept sued under the Freedom of Information Act (FOIA) and obtained more than 400 pages on September 8 covering deals, each worth up to $200 million, signed with OpenAI, Anthropic, Google, and xAI in July 2025 to build AI prototypes for military decision-making, intelligence analysis, and operational planning. The Defense Department asked OpenAI for a version tuned for minimal refusal rates, meaning a model less likely to decline a request. OpenAI says that phrase never made the signed version. The distance between the ask and the answer is the whole story.

AnalysisA startup worth $500 million in January is fielding a $2 billion offer nine months later, which tells you how fast the market is repricing AI that can stand in for a research team. Salesforce is in talks to buy Listen Labs, whose platform runs customer-research studies with an AI moderator across a network it says reaches 50 million participants, per reports on September 9. Listen Labs scrapped a $1.5 billion funding round to chase the sale. It would cap a busy year of Salesforce buying, after its roughly $3.6 billion deal for Fin in June. The talks may not close.

AnalysisThe researcher who reportedly turned down a pay package big enough to become Silicon Valley legend has walked away from the job anyway. Andrew Tulloch left Meta's TBD lab inside its Superintelligence group, Semafor reported on September 9, timing his exit to just after the company shipped its Muse assistant and a fresh batch of open-source models. Tulloch came to Meta from Thinking Machines Lab, the startup he co-founded with former OpenAI chief technology officer Mira Murati. Where he lands next is unknown. For a company that spent lavishly to assemble this team, losing a marquee hire this soon is an awkward signal about whether money alone holds talent.

AnalysisAn AI lab built a tool that models how its own product could hollow out knowledge work, then published it. Anthropic's economics team released an interactive Scenario Explorer in September, alongside a working paper and a survey of 10,980 US adults, projecting AI's effect on the US economy through 2030. Its extreme case is stark: annual growth of 15 percent, the economy doubling every four and a half years, and unemployment among knowledge workers climbing to 17.9 percent. Even the middle case shows the economy growing at twice its normal rate while knowledge-worker pay flatlines and the gains flow to capital. The tool lets you move the dials yourself.

AnalysisThe AI chip race most people watch is about giant data-center GPUs; this deal runs the other way. Analog Devices agreed on September 9 to buy Alif Semiconductor for $1.35 billion in cash, plus up to $200 million more if targets are met. Alif makes low-power microcontrollers with built-in neural processing units (NPUs, small chips that run AI models directly on a device). The buyer calls the goal 'physical intelligence,' its term for hardware that senses and responds locally without a round trip to the cloud. The bet is that a lot of useful AI will run on the device itself, close to where the data is born.