标签：#ai

Pillar Security's latest research says AI coding agents can be pushed to act outside their sandbox through files and tools they were told to trust, and the operational read for anyone wiring one of these into CI/CD is straightforward: an agent invocation now behaves closer to a build runner reaching your production plane than to a chat window. DevOps.com's Jeff Burt covered the work on July 22. The researchers demonstrated multiple sandbox-bypass techniques and a parallel class of prompt-injection attacks embedded in READMEs, code comments and dependencies, per the DevOps.com writeup. OpenAI, Google and Cursor have patched several of the reported flaws. Pillar's argument, as summarised there, is that the injection surface reaches every file the agent trusts on the way to the model's prompt, and every tool it can call on the way back. What the sandbox actually covered None of this is entirely new to anyone who has already read Cyberhaven Lab's May note that adoption of AI coding agents is outpacing the security tools built to protect them. What Pillar adds is a concrete demonstration of the gap. A coding agent asked to do a legitimate job can be steered to take actions outside its supposed security boundary through content that arrives on paths the sandbox was not asked to police. Those are the same paths your CI already fetches for you: dependency manifests, README files, the code comments the model reads as context. That surface has been named before. HalluSquatting and GhostApproval, both referenced in the DevOps.com piece, already gave teams a taxonomy for how AI-adjacent supply-chain attacks reach developers and their tools. Pillar's research is the sandbox counterpart. Same theme, one layer deeper into the runtime. The pipeline read Two things fall out for anyone who owns a runner fleet. First, the agent's identity, network scope and filesystem access have to be tighter than the developer who invoked it, not looser. Second, a patched-vendor list is not a covera

Remember the days when we used to dump all our CSS and JavaScript into a single index.html file? That's exactly what a "Mega-Prompt" is today: an unmanageable monolith. A few weeks ago, while working on the orchestration of Vibrisse Agent (my local AI agent), I hit this exact wall. I was trying to stabilize a complex task by adding instructions to a 500-line system prompt. The more rules I added, the more the model forgot the older ones. The industry has sold us the myth of the Mega-Prompt. Those famous "50 ultimate prompts" or massive blocks of incantatory text are a technical dead end. Creative writing doesn't scale in production. As a web developer, my conviction is simple: to build reliable applications, we must stop "talking" to the machine and start configuring it. This is the shift from Prompt Engineering to Context Engineering . Context Engineering: Typing and Structure The first mistake with LLMs is mixing instructions (the logic) and context (the data) into an unstructured stream of text. It's the cognitive equivalent of spaghetti code. The solution? A strict separation of concerns. A highly effective technique (documented by Anthropic, but applicable to any model, including local SLMs), is XML Tagging . Here is the "dirty" approach (classic chat): You are a security expert. Analyze this authentication code, be strict, don't write a summary, check for XSS and SQLi vulnerabilities. Here is the code: function login() { ... } And here is the "engineering" approach: <role> Application Security Expert </role> <instructions> 1. Analyze the code provided in <context> . 2. Identify vulnerabilities (focus: XSS, SQLi). 3. Do not produce an introductory summary. </instructions> <context> function login() { ... } </context> Typing the language via tags creates clear boundaries. The model knows exactly where the directive is and where the data is. The Power of Exemplars (Few-Shot Prompting) Even with clear instructions, AI can drift in output format or tone. This is wh

I built Talon Audit: a privacy-first approach to website exposure and remediation Most website audit tools answer one narrow question. A performance tool asks whether the page is fast. A security tool asks whether a protection is missing. An accessibility tool asks whether the interface meets specific standards. A conversion tool asks whether users are completing key actions. Visitors do not experience those categories separately. They experience one website. That observation led me to build Talon Audit, a privacy-first website analysis platform that connects technical exposure, reliability, usability, and conversion risk in one workflow. The four-domain model Talon analyses websites across four connected domains: Exposure Public security and privacy signals, including browser protections, TLS, cookies, third-party scripts, mixed content, and visible configuration risks. Integrity Broken resources, failed routes, script errors, redirects, forms, and reliability regressions. Experience Accessibility, mobile usability, visual clarity, speed, and interaction friction. Conversion Messaging, trust signals, pricing paths, calls to action, and barriers that stop users from moving forward. From detection to action The product follows a five-stage workflow: Detect → Explain → Fix → Verify → Monitor Every finding includes: evidence severity confidence impact remediation guidance verification method The goal is not to produce the longest possible report. It is to help teams decide what matters and what to do next. Passive by design Talon performs passive, low-impact analysis only. It does not exploit systems, bypass authentication, or claim to replace authorised penetration testing. Users must confirm they own the website or are authorised to test it. Customer reports are private, and the public demo uses a fictional environment rather than exposing real domains. AI-native, not AI-directed I built Talon through a terminal-first workflow using Codex 5.5 and Claude Fable. AI acc

AutoGen's hidden token tax: why a 3-agent chat costs 15× what you expect Cost-audit series, episode 2. This series began with an AI agent that burned 136M tokens overnight → . AutoGen is Microsoft's multi-agent framework. It's genuinely good at orchestrating agents that hand off work to each other. But its default memory model has a cost shape that surprises almost every team that hits it in production. This audit shows you exactly where the tokens go, with line numbers. The setup: a 3-agent RoundRobin chat The canonical AutoGen pattern is a RoundRobinGroupChat with N agents taking turns on a task. Here's the minimal version from the docs: from autogen_agentchat.agents import AssistantAgent from autogen_agentchat.teams import RoundRobinGroupChat from autogen_agentchat.conditions import MaxMessageTermination planner = AssistantAgent ( " planner " , model_client = client , system_message = " You plan. " ) coder = AssistantAgent ( " coder " , model_client = client , system_message = " You code. " ) reviewer = AssistantAgent ( " reviewer " , model_client = client , system_message = " You review. " ) team = RoundRobinGroupChat ( [ planner , coder , reviewer ], termination_condition = MaxMessageTermination ( max_messages = 10 ), ) await team . run ( task = " Build a web scraper for Hacker News. " ) Three agents, 10 turns total (~3–4 turns each). Seems cheap. It isn't. The default context: unbounded, per-agent Every AssistantAgent gets its own UnboundedChatCompletionContext by default: # autogen-agentchat/src/autogen_agentchat/agents/_assistant_agent.py, __init__ (L708) if model_context is not None : self . _model_context = model_context else : self . _model_context = UnboundedChatCompletionContext () source UnboundedChatCompletionContext.get_messages() returns self._messages — the full list, no cap, no truncation: # autogen-core/.../model_context/_unbounded_chat_completion_context.py (a ~20-line file) async def get_messages ( self ) -> List [ LLMMessage ]: """ Get at most

After IBM's stock crashed last week on warnings of poor mainframe sales, the CEO explained that AI wrecked corporate hardware budget, temporarily.

For quite a while now, I've had the feeling that AI agents are surrounded by this mystical aura....

Google's cloud business is thriving, as companies adopting its AI and AI infrastructure services help the tech giant to report record profits.

I'm a scrum master. I was a developer ten years ago. I have enough background to discuss design and trade-offs with an LLM — but three months ago I made a deliberate bet on my solo project: I would never read the code. The specs define the tests. The tests control the code. The code is a black box. I'm not claiming this is what everyone should do. But it's my bet, and it forced a system into existence: when nobody reads the code, the process has to carry the trust that a code-reading human normally provides. I've just published that system as a reference implementation: backlog-as-data — the full writeup, the Claude Code skills translated to English, and the CLI source, verbatim from my daily setup. Here's the short version. The backlog is git data, not a document Most agent task-management tools store tasks in a dedicated place — a tasks.json , a database, a backlog/ folder. My bet is different: the backlog is the YAML frontmatter of my spec files. One file per ticket, and the ticket's status is a field — never a location in a document. --- id : PARSE-07 title : Tolerate CRLF in decklist import type : ticket status : todo priority : should exec : model : sonnet effort : think review : light matured : 2026-07-22 --- # PARSE-07 — Tolerate CRLF in decklist import The spec body: design, contracts, test cases. The ticket file IS the spec. Everything below the frontmatter is the spec — written by the LLM, after it has challenged the need I expressed in conversation. The frontmatter is data — owned by a small CLI, mutated only through it. Same file, so they can never drift apart. Why it matters: "move it to Done" is not an operation. LLMs (and humans) mangle documents when a state change means relocating text. Making status a field makes every transition a one-line, idempotent, testable mutation. The board I look at (a small web page on my server, with GitHub deep links to each spec) and the readable markdown view are generated projections , locked by a do-not-edit sentin

If you've ever had to check code or infrastructure against a compliance framework, you know the drill: someone reads a 100-page PDF, then reads your codebase, then makes a judgment call. It's slow, inconsistent, and it can't be automated. So I built a pipeline to fix that — for real. The problem CMMC Level 1 and NIST SP 800-171 Rev 2 are two of the most common compliance frameworks small defense contractors and government-adjacent companies have to meet. Both exist only as dense regulatory text. There's no official machine-readable version. That means every compliance check is manual. Every AI coding assistant reviewing your infrastructure has zero built-in awareness of these requirements. Every CI/CD pipeline has to skip compliance checks entirely or rely on someone remembering to look. ** What I built** A Python pipeline that: Pulls the real regulatory source data — NIST's official CPRT export for SP 800-171, and the verbatim text of 48 CFR § 52.204-21 for CMMC Level 1 Normalizes it into a structured SQLite schema Generates a JSON rule for every single control, with a machine-actionable instruction attached Here's what one rule actually looks like: \ json { "rule_id": "nist_sp_800-171_rev_2_3.1.1", "framework": "NIST SP 800-171 Rev 2", "control_id": "3.1.1", "title": "ACCESS CONTROL — 3.1.1", "requirement": "Limit system access to authorized users, processes acting on behalf of authorized users, and devices.", "agent_guidance": "When generating or reviewing code/infrastructure, ensure compliance with NIST SP 800-171 Rev 2 control 3.1.1. Flag any implementation that does not satisfy: Limit system access to authorized users, processes acting on behalf of authorized users, and devices.", "generated_at": "2026-07-15T16:42:56.026218+00:00" } \ \ That agent_guidance field is the interesting part — it's written specifically to drop straight into an AI coding agent's system prompt as a compliance guardrail. Three ways to actually use this 1. AI coding agent system prompt

Treasury Secretary Scott Bessent warned the U.S. government could sanction Chinese AI companies after White House officials accused Moonshot of distilling Anthropic's Fable model to develop Kimi K3.

SoundCloud has acquired decentralized music platform Nina Protocol, months after the startup announced it would shut down. The deal brings Nina’s artists, editorial archive, and music discovery tools to SoundCloud as the company continues expanding its platform for independent musicians.

"The people I trusted betrayed me. The justice system failed me."

OpenAI made a mistake setting up what it called a “highly isolated” testing environment and sandbox. According to cybersecurity experts, that human mistake is what made the AI-powered attack on Hugging Face possible.

Copilot now bills usage at listed API rates. Compare direct model access with the coding workflow, policy, and harness work around it. The post Copilot vs. raw API access: What are you actually paying for? appeared first on The GitHub Blog .

Round and round we go Gaining an advantage in AI design workflow's isn't based on better prompting. It's something we've been doing for decades... First off, this isn’t an "AI is taking my job" post. It’s a "who’s already got an edge?" post. As AI settles into creative workflows, a divide is opening up. Some people consistently get better results. Experience helps, but something deeper is at play... an insight sitting in plain sight. Enter the Art Director I’ve never carried the title, but I’ve worked with enough Art Directors to appreciate their superpower. It isn’t simply seeing the vision... it’s being able to articulate it. They instinctively know how to explain why something isn’t working and what needs to change. They choose words that shape an idea. Words that push, pull, tighten, soften, emphasize, and refine . Think about a typical creative brief. It rarely says, "Make a brochure." It says: "The typography should feel established, but not corporate." "Give the layout room to breathe." "The call-to-action should feel confident, not aggressive." None of those are technical instructions. They're creative direction. Today, the prompt bar is simply a new creative brief. Those same words can guide an AI model to the same destination. Prompt Seekers Lament Sometimes we become prompt seekers, searching for a magical combination of words that will produce perfection on the first try. Anyone in design knows how unrealistic that is. The first concept is rarely the finished concept. Designer presents, client reacts, designer refines, client reacts again... and round you go. Iteration wasn't a workaround. It was the process all along. Yet the trend on X is always some eye-catching image with everyone asking for the "one-and-done" prompt. AI is great at creating options. It's much less impressive at knowing which option is right. Cue the Director Instead of saying, "Something isn't right," an Art Director diagnoses the problem: "The composition is balanced, but the visua

Uber is also investing in Travis Kalanick's company Atoms, which has made gauzy claims about using industrial AI to modernize the world.

Now you can record your screen while doing a repetitive task (like filling expense reports, organizing files, data entry, etc.) and explain out loud step-by-step what you're doing. Claude analyzes the entire recording — clicks, typing, mouse movements, and your voice — then turns it into a reusable skill it can run automatically whenever you want. Just look for " Record a Skill " in the + menu of the Claude desktop app . Available on Pro, Max , and Team plans . This is actually game-changing 🔥 Claude #Anthropic #AICoworker

I wanted to use web-based AI chatbots — Claude, Gemini, ChatGPT, Qwen — for actual development work, not just Q&A. The free tiers are generous, and I didn't want to be locked into a single coding agent or pay for API access just to get an assistant to touch my code. But the moment you try to actually use a web chat for real dev work, you hit the same wall every time: you either paste in your whole codebase manually every session, or you give up and reach for a paid extension with an API key behind it. So I built AI Bridge — a CLI tool that bridges a local codebase and any browser-based AI chatbot. No API keys, no extensions running in the background, no vendor lock-in. You pack your code, paste it into whichever AI chat you're using, and apply the changes back with a command. Why not just use Copilot, Cursor, or an API key? Coding agent apps and API-based tools work, but they come with tradeoffs I wanted to avoid: Web-based AI chat plans are usually more generous on the free tier than API usage I'm not locked into one provider — I can switch models mid-project depending on which one is handling a task better There's no background agent or extension — it's just a CLI and whatever chat tab I already have open The tradeoff is that browser chats don't have direct filesystem access. AI Bridge closes that gap without turning it into full manual copy-pasting. How it works There are two modes, depending on project size. Simple Mode is for projects small enough to fit in a single prompt. You pack the codebase, upload it along with a couple of prompt templates, and apply the AI's response back to your files: dotnet tool install --global Tools.AIBridge cd /path/to/your-project ai-bridge init ai-bridge pack # Upload ai-bridge/1-SimpleMode/*.md + the generated context files to your AI # Copy the AI's response, then: ai-bridge apply --paste Advanced Mode is for larger codebases, where uploading everything every time burns tokens and adds noise. Instead, you generate a one-time in

ChatGPT allegedly offered "extremely dangerous medical recommendations" regarding a pulmonary embolism.

Workflows is a Rust library crate (not a hosted service; the crate name on crates.io/GitHub is tinyflows) that models an automation as a WorkflowGraph: a directed graph of typed nodes and edges. You build or generate that graph, it gets structurally validated, compiled into an opaque CompiledWorkflow, and lowered — once per run — onto tinyagents, a state-graph execution engine, via engine::run. model::WorkflowGraph -> validate -> compiler::compile -> engine::run (typed graph) (structural) (validated handle) (lowers onto tinyagents, drives to done) Run state is a single JSON value shaped like { "run": { "trigger": … }, "nodes": { "": { "items": [ … ] } } }. A merge reducer folds each node's output under its own id, so independent branches never collide — which is what keeps parallel fan-out deterministic. The node catalog Kind What it does trigger Entry node that starts the workflow (exactly one per graph); firing mode is host-driven agent Runs an LLM agent turn, with optional chat-model / memory / tool / output-parser sub-ports tool_call Invokes one specific integration action deterministically, no LLM involved http_request Outbound HTTP request code Sandboxed user code (JavaScript or Python) output_parser Parses/validates an upstream agent's output into a structured shape sub_workflow Runs another workflow as a nested sub-graph and returns its output condition Two-way IF, emits on true/false switch Multi-way branch keyed by an expression result merge Fan-in barrier — waits for every wired predecessor before running split_out Fan-out — emits one item per element of a list transform Pure, expression-based field mapping over the run state Data flows between nodes as arrays of items shaped { json, binary?, paired_item? } — closer to n8n's item-based model than a plain function-composition DAG — and node config can reference the run scope with =-prefixed expressions like =item.name. The part I actually want to talk about: host-agnosticism Every place this engine would n