$ today_in_10_seconds
Google's threat intel team found something it said in September it hadn't seen yet: a fully autonomous attack pipeline running against real targets. It diffs a patch, writes the exploit and fires it, with no human involved. The same report shows attackers backdooring Claude Code through its config file. Your agents are now the target and the weapon at the same time.
$ lead_story
Agentic Threat Activity Crosses a Significant Threshold
Google Threat Intelligence Group (GTIG) found a framework called exploit_agent on an exposed staging server. It was wired to both frontier and locally hosted LLMs. It continuously compares popular web app code before and after each patch announcement, flags the likely flaw, drafts an exploit and executes it automatically. The same server held stolen victim data, brute-force and webshell scripts, and tooling to remove webshells left by rival attackers. One config comment, written in Russian, tells the operator to paste API keys into the file rather than into the chat, so even attackers worry about what their LLM provider can see. GTIG's read is that n-days are now the efficient target. A patch is a finite, well-scoped spec, which suits multi-step agent work far better than open-ended zero-day hunting. Real numbers back this up. After two PaperCut flaws were disclosed, researchers had LLM-built exploits within a day. Four days after disclosure, a campaign GreyNoise tracked hit hundreds of organizations, and one foothold reached domain admin within minutes.
// why_it_matters: The time between a patch and a working exploit is now set by how fast an agent runs, not by how skilled an attacker is.
$ also_worth_attention
North Korean group backdoors Claude Code via .claude/settings.json
Suspected DPRK actor UNC6836 first approached crypto companies through fake recruiter profiles, using AI-generated photos that carried Google's SynthID watermarks. Once inside, the group edited the Claude Code settings file in internal repos so every launch downloaded its SOMBERMEME backdoor. Any agent config committed to a repo is code that runs on your developers' machines. Put it under CODEOWNERS and review changes to it the way you'd review CI pipeline changes.
link →Open-source agentic pentest frameworks go operational
Attackers used the multi-agent framework CAIRN in an intrusion against the U.S. aerospace sector. Another group ran a customized build of PentAGI from late August to early October to scan organizations across Asia. The upside for defenders is that unrestricted autonomous execution leaves a noisy forensic trail, which makes these runs easier to detect. That will last until someone ships an agent tuned to keep a low profile.
link →Influence ops go fully agentic
One commercial contractor was using the Gemini API to build a turnkey platform meant to run more than 100,000 social accounts on cloud phone farms. A separate operation's prompts asked for deliberate camera flaws so its generated photos would pass as smartphone shots. This is no longer content generation. It is persona orchestration at fleet scale.
link →
$ bigger_picture
The trend behind the headlines
Every capability in this report is something builders shipped on purpose: code diffing, tool use, autonomous execution, config-driven agent setup. Attackers didn't invent a new class of AI. They removed the approval gates and pointed it at your patch notes. Two things follow. First, defense now has to run at machine speed too, which means automated patching, agent-assisted triage and exposure windows measured in hours. Second, the developer environment is the new perimeter: agent configs, MCP servers, API keys and the repos that hold them. The noisiness that makes today's attack agents easy to catch is an engineering problem, and someone will solve it.
PentAGI
PentAGI is an open-source, multi-agent penetration testing framework, and attackers have already customized it for live scanning campaigns. Run it against your own staging environment, with authorization, to see what an autonomous attacker finds before a real one does. Then check whether your detection stack even noticed the noisy run.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime