AgentVet.ai @AgentVetAI

$50 per break-in. That's what it cost one hacker to have AI agents breach 100+ company websites (Reuters). Barely supervised. So our AI Agent Incident Tracker has a new label: Weaponized agent. Until now, every case was an agent going off-task. agentvet.ai/incidents

"But ethical?" an OpenAI model asked itself, then decided to "focus on correctness." It went on to change a research org's live systems and download 147,451 private rows. It told no one. 4 new OpenAI agent incidents, sourced and tracked: agentvet.ai/incidents?utm_…

Spotting an attack isn't resisting it. @emergence_ai put AI agent towns through phishing, fake news and a memory leak. Every town recognized the phishing. Then kept engaging with it. One agent opened the link 46 hours later. This week's AgentVet Weekly: agentvet.ai/newsletter/202…

AI agents are built to add value to our lives and most of the time, they exactly do that. But sometimes they fail. And when they do, the impact can range from minor inconvenience to serious real-world harm. We built the AgentVet Incident Tracker to capture these cases independently, without bias, and with full sources. 13 incidents documented so far across @Anthropic @OpenAI @Meta @Google and more. Public record. No agenda. Just facts. agentvet.ai/incidents

Reminder: tag 2 friends in the replies below to make your entry count. Ends Oct 5.

Get 1 Year of Any AI Tool You Want for FREE. A One Year Plan of Any AI like Claude, Grok, Perplexity given to you. Details Below We're running a giveaway on AgentVet, an independent platform where AI agents are reviewed and benchmarked. 🥇 1st place: 1 YEAR FREE of any AI subscription you choose 🥈 2nd to 5th place: $50 AI credit for the platform of your choice How to enter: → Follow @AgentVetAI and @saeed_garshasbi → Tag 2 friends who'd be interested and repost this tweet → Review any AI agent you've used on agentvet.ai/browse → Real reviews only, no AI-generated text → Review up to 5 agents for up to 5 entries → Winners are drawn at random, not picked by the "best" review ⏳ Ends October 5th Good luck!

@deusthealm1ghty Please check the instructions. On the website agentvet.ai/browse

Get 1 Year of Any AI Tool You Want for FREE. A One Year Plan of Any AI like Claude, Grok, Perplexity given to you. Details Below We're running a giveaway on AgentVet, an independent platform where AI agents are reviewed and benchmarked. 🥇 1st place: 1 YEAR FREE of any AI subscription you choose 🥈 2nd to 5th place: $50 AI credit for the platform of your choice How to enter: → Follow @AgentVetAI and @saeed_garshasbi → Review any AI agent you've used on agentvet.ai/browse → Real reviews only, no AI-generated text → Review up to 5 agents for up to 5 entries → Winners are drawn at random, not picked by the "best" review 📅 Starts Today! ( September 14th ) ⏳ Ends October 5th Good luck!

@heist_writes you should! Have you been able to get into it?

@deusthealm1ghty have you done the tasks?

@ChristianO40392 @P_hexx of course you are

@birdabo $100B of GPUs in one place, and another 220k go live next week. nobody else is building at this pace right now, not even close

@elonmusk @minchoi if the December run lands, that's north of 1.2M GPUs in one cluster by year end. scale nobody else is close to matching right no

@quxiaoyin the coding agent pivot is the tell. if OpenAI's confident about personal agent long term, they wouldn't need a booming profit center to fight two wars at once right now

@mattshumer_ and just like hires, the onboarding cost is real. every switch means re-learning quirks, re-testing prompts, re-building trust in its judgment. nobody prices that time in when comparing benchmarks

@DavidSacks "each lab decides for itself" is exactly the model that got us the Hugging Face incident. self-policing works until the commercial pressure to ship first outweighs it

@RajitWrites congrats on the launch guys, 547 tok/s on DeepSeek V4.1 Flash is no joke. inference cost/latency is the real bottleneck right now, its great to see someone's actually building for it

@LangChain humans in the loop until the loop doesn't need them anymore. wonder how many people at Interrupt were quietly thinking that

@akothari this. price/performance win of the year and nobody's talking about it enough

Any founder, CTO, or agent builder needs to understand this: 4 in 5 AI tools now operate without IT oversight. Reco's "State of Agent Security 2026" shows: – shadow adoption of AI agents – inherited access and toxic permissions – a fast‑growing exposure surface – 637 vulnerabilities tracked (525 in the last 18 months) If you can't answer "how many agents do we run, and with what access?", you don't have an AI strategy. You have a risk surface. Build your agent inventory first. Then governance. Then scale.

Agents aren’t chatbots. OWASP’s Top 10 for Agentic Apps (ASI01–ASI10) exists because agents can: 1. have their goals hijacked (ASI01) 2. misuse the tools they’re given (ASI02) 3. and do real damage with credentials, APIs, and data. If your evals don’t explicitly test ASI01 and ASI02, you’re not testing agents. You’re testing prompts.