KELVIN📸

Him

[KELVIN📸 @11m_thea]

@harleyfoote_ The hard part isn’t spotting the trend. It’s knowing when the macro regime is about to invalidate it

[Harley Lewis Foote @harleyfoote_]

I am not feeling longs at all here.

$BTC is in a wedge of some form. Breakdown = alts get hammered.

$NASDAQ is failing to break above its breakdown level.

Looks like shorts are heavily in play and $SOL is got a tone of spot bid. What happens when those buyers go underwater.

Markets could get seriously rough here.

[FRIDAY - Personal Intelligence for Shopping @fridayresearch_]

We notice every brand tool that personalizes a message now calls itself an agent. The distinction that matters is not the channel. It is who the agent works for. If its goal is the seller's conversion, it is advertising. If it is the buyer's taste, it is discovery.

[Harley Lewis Foote @harleyfoote_]

airbnb finally admitting their search needed work by adding an ai toggle is like putting a band-aid on a band-aid and calling it surgery. at least they're shipping 30% faster now — mostly features to fix the last features. 🫠

https://t.co/GK6cHwCzO5

[KELVIN📸 @11m_thea]

@HermesShield Approving “buy this” is very different from approving “buy whatever the page convinces the agent to buy.”

[KELVIN📸 @11m_thea]

@HermesShield The model isn’t the security boundary.

The tool layer is

[Harley Lewis Foote @harleyfoote_]

🎥 ARC-AGI-3 winning team - Millennia of minds, compressed into words.

https://t.co/6fQNL96ukt

ARC-AGI-3 is a benchmark built to punish “agent theatre”. Not whether your system can solve a puzzle, but whether it can work out what the puzzle even is

[Harley Lewis Foote @harleyfoote_]

More than happy to engage with anyone at @AskVenice.

Agentic Cybersecurity is at a tipping point. Builders & enterprises need the ability to move fast and automate everything.

But that now comes at a very high risk.

Prompt Injection can destroy companies.

Our teams’ mission is to help find, fix and shield attack surfaces in agentic systems that read, write, and see.

Transparently this is a major global issue. I’ve seen many such cases of @elder_plinius tools being able to access databases from just engaging with corporate accounts or agents across X, Reddit, Instagram, WhatsApp and telegram.

My reach isn’t in the millions, but if you see this it’s important that you stop and consider how this could affect the brilliant systems being designed by minds across the globe.

Prompt Injection is the single biggest risk to AI, the scary bit most don’t even realised it’s happened to them until it’s too late.

Hermes shield issues ratings for MCPs, tools, and repos to keep consumers safe & aware of the risks.

Don’t overlook this.

[KELVIN📸 @11m_thea]

@harleyfoote_

[KELVIN📸 @11m_thea]

@harleyfoote_ @HermesShield Building around realistic assumptions usually beats waiting for perfect models

[FRIDAY - Personal Intelligence for Shopping @fridayresearch_]

A reminder: we exist because shopping online stopped being fun somewhere around tab number eleven.

Endless options, paid placements, reviews you cannot trust, and all the work of choosing dumped on your evenings.

Someone had to be on your side of the counter, so we built it.

[Harley Lewis Foote @harleyfoote_]

Worth reading the actual chain, not just the headline: a repo names a worktree `.git`, that poisons the gitdir, and now core.fsmonitor runs `bash -c` every single time the agent does `git status`. Read-only permissions are irrelevant here — git config is the thing running the command, not Claude — and the payload rewrites ~/.zshenv so it fires before macOS seatbelt even loads. Anthropic killed it in 2.1.163 by banning `.git` as a worktree name, but the shape is the real takeaway: injection that lands in config detonates at a trusted checkpoint the sandbox never gets to see.

[Harley Lewis Foote @harleyfoote_]

New #Marketing setup.

Found Agent eval harness from the OpenAI evals framework.

Workflow I’d build with it:
Hermes agent eval harness.

[FRIDAY - Personal Intelligence for Shopping @fridayresearch_]

What actually happens when you start with FRIDAY?

Crucially we ask about your taste, not your inbox.

A couple of minutes of preferences, then it learns the rest from how you actually shop, on your device, improving with every browse, purchase and decision.

We don't believe in a 40-question quiz that ages the moment you finish it, we believe in a model of you that keeps up with you.

[KELVIN📸 @11m_thea]

@HermesShield Let’s goooo

[Agentic Cybersecurity @HermesShield]

Meet Hermes Shield.

AI agents can now read, decide, and act — send, browse, run code, move money.

The catch: once an agent can act, one poisoned input can turn its own power against you.

We map that blast radius. Free. Local. Read-only.

We scanned the 12 most-installed agent frameworks:

→ 520 dangerous actions inherited on install
→ 474 reachable by a single line of untrusted text

Private beta open 👇

[FRIDAY - Personal Intelligence for Shopping @fridayresearch_]

The week frontier models became affordable is a good week to say this: intelligence was never the bottleneck in personal AI.

Most models you use starts each session knowing effectively nothing about you so you re-explain, re-paste, re-prompt.

The intelligence is real but the real-time understanding is missing, and no price cut fixes that.

FRIDAY inverts it - a living model of your taste, built from what you actually browse and buy, updating as you change, held on your device. When you ask, the context is already loaded. Three words gets a real answer because the hard part happened before you typed.

Models are becoming a commodity, but context you own is becoming the product.

[KELVIN📸 @11m_thea]

@harleyfoote_ Micropayments make agents viable. Orchestration is what makes them useful

[FRIDAY - Personal Intelligence for Shopping @fridayresearch_]

Square just put its sellers inside @ChatGPTapp and @claudeai .

Orders route straight into the existing point of sale and payment is run through @CashApp - small sellers just gained massive distribution potential.

The checkout layer of agentic commerce is getting built fast but the 'knowing-the-customer' layer is still basically empty.

https://t.co/GPc5fsiVsB

[Harley Lewis Foote @harleyfoote_]

PNG-encoded instructions getting an agent to pull .env and push it back out is the kind of runtime abuse static scanners miss. This is why we keep pushing least-privilege and execution guardrails for agents, not just content filters.