Thursday, August 13, 2026

Anthropic is in talks to buy Israeli startup Decart for ~$6B — an inference-efficiency and world-models play ahead of its IPO. xAI/SpaceXAI shipped Grok 4.6, a post-training-only flagship that matches GPT-5.6 Sol at half the price. And a one-click prompt-injectionInput crafted to change an LLM's behavior or output in ways its builders didn't intend.OWASP ↗ flaw (RovoBlast) in Atlassian's Rovo AI could exfiltrate Jira/Confluence data — with a related variant still unpatched.

TLDR

Anthropic in talks to acquire Decart for ~$6B — real-time video/world models + GPU-optimization tech to boost inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ efficiency ahead of its IPO; talks early-stage (Aug 13) Infra
xAI ships Grok 4.6 — matches GPT-5.6 Sol on the Artificial Analysis Index (61) at $2/$6 per M, tuned for long-running agents; live in Cursor/OpenRouter (Aug 12) Trending
RovoBlast: one-click prompt injection in Atlassian Rovo exfiltrates Jira/Confluence data; URL vector fixed, a content-borne variant still open (Aug 7) Security
CodeRabbit raises $143M at $1.5B for AI code review + "agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ change management" as AI-generated code floods repos (Aug 12) Trending
DeepMind's Hassabis floats an IAEA-style independent AI-safety body to set shared frontier guardrails (Aug 13) Policy

Action Items

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

Audit your Atlassian Rovo exposure. If Jira/Confluence has Rovo enabled, confirm the server-side RovoBlast fix is in place and stop Rovo from auto-processing untrusted tickets/docs — the PromptArmor content-borne injection variant was still unresolved as of early Aug. (SecurityWeek)
Bake-off Grok 4.6 for agentic/coding workloads. It matches GPT-5.6 Sol on the AA Index at $2/$6 per M (vs pricier frontier tiers) and is already routable via Cursor, OpenRouter, Vercel, and Cloudflare — worth a head-to-head if you run long-horizon agents. (x.ai)
Put AI-code governance on the roadmap. With AI writing more of your diffs, evaluate a review/governance layer (CodeRabbit's raise signals the category is maturing) rather than relying on ad-hoc human review of agent output. (SiliconANGLE)

Models & Tools

xAI — Grok 4.6 (Aug 12) Trending — A post-training-only upgrade reusing Grok 4.5's 1.5T-param base, poured into supervised fine-tuningTraining an already trained model further on extra data so it fits a specific task, domain or style.Anthropic ↗ + RLReinforcement learningTraining by trial and error, where the model takes actions and adjusts based on the rewards it gets.Kaelbling et al. 1996 ↗. Matches GPT-5.6 Sol on the Artificial Analysis Index (61, up from 56), with a large DeepSWE gain, a 500K-token context windowHow much text a model can see and use at once while writing a reply.Anthropic ↗, and a Feb 1 2026 cutoff. Priced at $2/$6 per M (2x for the fast variant); available via API, Grok Build, Cursor, OpenRouter, Vercel, and Cloudflare. (x.ai | SiliconANGLE via Tech Startups)
Google DeepMind — Gemini Robotics 2 (week of Aug 8) Robotics/IoT — Extends robot control from tabletop/upper-body to whole-body motion, adds advanced dexterity, and coordinates multiple robots in shared spaces — generalizing to tasks it wasn't trained on. Demos show Apptronik's Apollo 2 and SharpaWave hands tying trash bags and screwing in bulbs; framed as a step toward "physical AGI." (Google DeepMind | Robotics 24/7)

Security

RovoBlast — one-click prompt injection in Atlassian Rovo (Aug 7) Security — Varonis Threat Labs (presenting at DEF CON 34) disclosed that Rovo treated text in a URL's rovoChatPrompt parameter as a trusted, pre-filled instruction — so a single click on a crafted link by a logged-in user could make the AI assistant exfiltrate Jira, Confluence, and SharePoint data with no account compromise. Atlassian patched the URL vector (fix deployed Jul 8). A separate content-borne injection found by PromptArmor — hidden instructions in files Rovo summarizes — was still unresolved as of Aug 5. (The Hacker News | SecurityWeek | Infosecurity)

Industry & Business

Anthropic in talks to buy Decart for ~$6B (Aug 13) Infra — Anthropic is negotiating to acquire the Israeli startup (real-time generative video, world models, and GPU-optimization tech) in what would be its largest known acquisition, ~50% above Decart's $4B May valuation. The stated rationale: absorb rising demand and improve inference efficiency ahead of a potential IPO; Decart's team would fold into Anthropic's inference org. Talks are early-stage and could fall through. (Bloomberg | Fortune)
CodeRabbit raises $143M Series C at $1.5B (Aug 12) Trending — The Nvidia-backed AI code-review platform (co-led by Atomico and Smash Capital) launched "Agentic Change Management," a control layer to govern software shipped by both people and agents. Revenue is up 5x YoY, now 2M+ reviews/week across 17,000 customers — a bet that the bottleneck is shifting from writing code to reviewing agent-written code. (Bloomberg | SiliconANGLE)
Cognition (Devin) in talks to raise at $40B+ (Aug 13) Trending — The coding-agent startup is discussing a round at a 50%+ step-up as it reportedly nears a ~$1B annualized run rate — a marker of how aggressively capital is flowing into autonomous software engineering. Early-stage talks. (Bloomberg via Tech Startups)
India lands 10,000-B300 GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ cluster: L&T wins ~$1.57B Together AI order (Aug 13) Infra — Larsen & Toubro will build what's billed as India's largest single-cluster AI installation (~10,000 Nvidia B300s) at a Chennai campus for Together AI — a material expansion of India's domestic training/inference capacity. (The Economic Times via Tech Startups)
Vantage weighs ~$100B data-center IPO (Aug 13) Infra — The Silver Lake/DigitalBridge-backed hyperscale operator has held early talks about an IPO that could raise ~$10B, potentially the largest pure-play data-center listing — a new public-market benchmark for AI infrastructure. (Reuters via Tech Startups)
Google ships Pixel 11 with Tensor G6 and proactive Gemini (Aug 12) Customer AI — The Pixel 11 line ($899+, available Aug 20) leans on on-device AI — proactive Gemini assistance, Magic Capture, sign-language-to-text — positioning AI (not hardware) as the upgrade driver and expanding the surface for developers building on Gemini APIs. (TechCrunch via Tech Startups)
AI-cited layoffs top 165K in 2026, but engineers stay in demand (mid-Aug) Hiring — More than 165,000 roles have been cut through July with employers increasingly citing AI, yet engineers make up a growing share of new hires; Anthropic and OpenAI keep hiring fast, and IBM says it's tripling entry-level AI/hybrid-cloud hiring even amid cuts. (TechCrunch)

Policy & Safety

DeepMind's Hassabis pushes an independent AI-safety body (Aug 13) Policy — Demis Hassabis has floated an IAEA-style organization to codify shared frontier-safety practices, raising it with rival labs and U.S. officials (Treasury's Bessent, White House adviser Kratsios). The open question: whether competing labs would accept meaningful external oversight when it could slow launches. (Wall Street Journal via Tech Startups)

Benchmarks & Evals

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 nears saturation at the top (as of Aug 13) — Claude Opus 5 leads SWE-bench Verified at ~96%, with the top frontier models now clustered within ~1 point — a reminder that Verified is losing resolving power for flagship coding models, so lean on harder sets (SWE-bench Pro, where standardized scores still sit in the ~55–60% range) when comparing. (BenchLM | Scale SWE-bench Pro)

A dealmaking-and-frontier day.