This Google DeepMind’s paper is a serious warning for anyone using autonomous agents today.

Gives the first clear taxonomy of 6 attack types where harmful websites can detect AI agents and show them hidden content humans never see, like

- Instructions buried in HTML comments or white-on-white text

- Steganography in image pixels

- Override commands in PDFs, metadata, or even speaker notes

- Memory poisoning that persists across sessions

- Goal hijacking and cross-agent cascades in multi-agent setups

The real security problem for AI agents is not just the model, but the environment it reads.

The web itself can be weaponized against autonomous AI agents. As agents increasingly browse the internet, read emails, execute transactions, and spawn sub-agents, the information environment becomes an attack surface.

In one cited benchmark, hidden prompt injections embedded in web content partially commandeered agents in up to 86% of scenarios, sub-agent hijacking working 58–90% of the time, and data exfiltration attacks clearing 80% across five different agent architectures.

That reframes the whole debate.

We usually talk about model safety as if the danger sits inside the weights, but agents do something more fragile: they browse, retrieve, remember, and act on untrusted material in real time.

Here’s the thing to worry about.

A web page does not have to look malicious to be dangerous to an agent, because the agent may parse what humans never see: hidden HTML comments, metadata, CSS-hidden text, formatting syntax, or adversarial content embedded in images and other media.

The threat gets more serious once memory enters the loop.

If an agent uses RAG or persistent memory, poisoning no longer has to win in one shot. It can sit quietly in a corpus or memory store and activate later, which is why the paper highlights results showing latent memory poisoning above 80% attack success with less than 0.1% data contamination.

Starlink V3 is about to be INSANE.

V3 is bringing 100X more bandwidth and sub-5ms latency from ORBIT.

Lower flight, more satellites, more power.

Internet beamed from space is about to humble every provider on Earth.

Anthropic CEO:

“At Anthropic, we essentially have Claude designing the next version of Claude itself.”

Meanwhile, Anthropic says every new model is given code that trains a small AI model and asked to make it faster.

Claude Opus 4 achieved a 3x speedup. Mythos Preview reached 52x.

Recursive self-improvement is starting to look real

The models are helping build what comes next.

The narrative that AI will wipe out enterprise SaaS overnight is one of the most misunderstood ideas circulating in markets right now, and the evidence does not support it (Save this).

#DavidSacks made this case directly and the logic is worth working through carefully.

Salesforce is a system of record debugged by millions of customer support tickets over twenty five years, stress tested across thousands of enterprise deployments and deeply embedded into revenue operations at the largest companies on earth.

The idea that a CFO will replace that with probabilistically generated code from an AI assistant without compliance guarantees, integration depth, audit trails, and enterprise support infrastructure is not how these decisions actually get made.

The market has been pricing in the existential version of this risk anyway and the results have been extreme.

Over $1 trillion in SaaS market cap was erased in the first week of February 2026 alone.

Global SaaS spending is still projected to grow from $318 billion in 2025 to $512 billion in 2028 which is not the trajectory of a category being killed.

The operating reality is entirely disconnected from the stock price narrative.

ServiceNow beat earnings nine consecutive quarters in a row and its stock crashed 11% on the same day.

Salesforce raised its full year forecast to $41.5 billion on record results and the stock still fell.

Sacks makes an important distinction between survivability risk and value capture risk.

The survivability risk, enterprises ripping out Salesforce for AI generated software is largely overstated.

The SaaS products genuinely at risk are narrow ones charging high prices for underused features with no proprietary data and low switching costs.

The value capture risk is real and it is the more sophisticated threat.

AI orchestration layers like Claude CoWork are being designed to sit above all of these tools pulling data from Salesforce, ServiceNow, and Snowflake simultaneously and owning the user's primary workspace in the process.

If enterprise users move from living inside Salesforce to living inside an AI agent that calls into those systems on their behalf, the SaaS platforms do not disappear but rather become infrastructure.

The expansion revenue, the premium pricing power and the next decade of value creation all migrate to whoever owns that orchestration layer.

AI agents are not killing SaaS!

But UI-only software may not be enough in an AI world.

Jensen Huang (CEO $NVDA) pointed out that AI agents still need software to pull data, trigger workflows, update records, and complete real work.

The key is that software needs to become readable and usable by agents.

Clean APIs, structured data, workflow access, permissions, etc.

Because agents will not rebuild decade-long, proven software stacks from scratch, especially in a world where compute is that expensive.

It is much easier to navigate inside existing tools than recreate them.

It is also easier for customers.

No painful migration, no huge change-management process, no need to rip out systems that already run the business.

That is good for platforms agents can work inside: $NOW, $MSFT, $PLTR, $SNOW, …

In the agent era, the best SaaS companies may be the ones agents depend on in the background!

🏆Wow! Nemotron 3 Ultra beat Qwen3.5-397b (and 27b) in with a score of 47.7 versus Qwen3.6-27B 45.8 on #ArtificialAnlys. It is big. Going to need 256GB shared memory to run it locally at least.

Unsloth Quants
» UD-Q4_K_S = 328GB
» UD-IQ3_XXS = 223GB
» UD-IQ2_XXS = 194GB

artificialanalysis.ai

NVIDIA DOESN'T JUST MAKE THE CHIPS ANYMORE. THEY'RE BUILDING THE MODELS TOO.

Nvidia just shipped Nemotron 3 Ultra - a 550B MoE open model built specifically for long running agents.

5x faster inference, 30% cheaper than other open frontier models on complex agentic tasks.

Nemotron 3 Ultra is fast and genuinely good

Compared it with 3 frontier models: DeepSeek V4, MiniMax M3, and Qwen 3.7 Max on 2 prompts

very impressive results

This robotic hands will cuase some layoff in massage parlors 😅 Co-ordinated finger movements. Fist clenching, pointing & spreading. Complete hand closures. Palm opening and precise pinching actions and digit control. Xynova at ICRA 2026 in Vienna.

Boston Dynamics electric Atlas just landed a ghost rabona, a behind-the-leg trick kick that takes full-body balance most people never learn.

It is the newest drill in the Hyundai football series teaching Atlas athletic movement.

How long before a humanoid out-dribbles a pro?

Wuji Hand 2 Makes Its Global Debut at ICRA 2026

Next-generation dexterous hand with 20 DOF, backdrivable actuation, and improved thermal design for sustained operation.
Built for real-world dexterous manipulation.

Starlink V3 satellites

Bandwidth per Satellite:
• V2: 96 Gbps
• V3: 1,024 Gbps

Bandwidth per Launch:
V2: 2,600 Gbps
V3: 61,000 Gbps

Deployment per Launch:
• V2: 27 satellites (on Falcon 9)
• V3: 60 (on Starship)

Starlink V3 satellites will begin to be deployed in late 2026 on Starship.

Perhaps the coolest slide from the presentation: