What a crazy week if you are in security - plex2shell: PoC dropped oct 16, exploited in the wild oct 17 - beyondtrust: remote support zero-day used since aug 29, china-nexus actor, now in CISA KEV - ivanti connect secure: auth bypass zero-day exploited since oct 3, also in KEV - next.js: two critical unauthenticated RCEs, upgrade to 16.4.1 or 15.5.28 - litellm: v1.107.0 reported backdoored, steals cloud, ssh and ai creds. pin 1.106.2 and rotate secrets - emudhra: a certificate authority’s keys were crackable, update your browsers and trust stores - tidb cloud: pingcap breached through stolen credentials - php: two critical unauthenticated RCEs, already exploited attackers go from PoC to payload in a day, patch cycles take weeks and coding agents now hold your cloud keys, ssh keys and a shell forever Patch fast and check every tool call and every install before it runs we open sourced that part: github.com/PrismorSec/prismo…

Claude auto just killed our startup, because the auto-mode now asks users before deleting anything? It's great anthropic is focusing on model alignment, but we are building something bigger Harness alignment agonstic of agent and LLM We ran various dangerous commands and prompt injections against vanilla claude code and then with Prismor and the harm rate dropped from 68% to 0% there are lot of other unsolved business specific cases where users would like to control their fleet of agents on: •⁠ ⁠egress rules •⁠ ⁠supply chain package rules •⁠ ⁠secret redaction •⁠ ⁠custom policies which is why we are building Prismor, completely open source and independent of models github.com/PrismorSec/prismo…

Today we're open sourcing Patchbot the tool we've been using internally for a while to find and patch vulnerable dependencies It scans your lockfiles against OSV.dev or your own threat feed, bumps the version itself and (optionally) calls a coding agent when the bump breaks the build the tool has helped 8 enterprises and 50+ repos patch 1500+ vulnerable dependencies pip install patchbot github.com/PrismorSec/patchb…

Found in our data: open internet tool calls are where agents get compromised. OpenAI came to the same conclusion independently The most capable AI lab on earth answered a cyber incident by isolating its models from the internet

You get it, its a tough gap to resolve I feel it is model agnostic problem, it requires scrutinizing what the code does when it runs Prismor doesnt try to detect the intent of generated code, it just gates the agent's actual tool calls like file writes / network egress at run time So the actions are caught right there, weights are not a concern then

Recently, Anthropic pushed Inference Hooks Claude Enterprise orgs can now route every prompt on Claude app, Claude Code, and Claude Cowork through their own AI security server before the model responds You shouldn't need a different security policy for: - someone chatting with Claude web app - someone running Claude Code - someone using Cowork Prismor now directly plugs into Claude's Inference Hooks The hook receives Anthropic's signed request, runs it through Prismor's existing policy engine and returns an allow/deny decision within the hook's timeout So the same policies that govern agent tool calls on developer machines can now govern what reaches Claude itself It's open source and live today and works with your Claude Enterprise plan pip install prismor prismor inference-hook serve Docs: prismor.dev/docs/claude-infe…

AI is accelerating traditional cybersecurity, but who secures the agents? Agent adoption is increasing 42% YoY, be it robotics, enterprises, everywhere! therefore we’re building an independent open source platform to stop agents from going rogue Thanks to @aihouseseattle , @Taylor_Soper and Sri last week for giving us the opportunity, we had a great time to present the vision and traction of Prismor developers can get all agent traces and harmful actions with just one command - all opensource

So I had a fun trip to vegas last week at Blackhat looking for who is solving the security gap in Agents Funny enough, came back with a longer list of questions than answers The question I went with: what is everyone actually doing about security for AI? Met a fair number of CISOs and vendors and it all felt polished until I realized they're talking about different things and definitions AI Security which is AI doing the security work. Feed it SIEM data, it triages, prioritises, auto-fixes alerts, where AI is the tool, soc stuff etc Security for AI which is about securing the agents themselves. Identity, access, governance, runtime control, audit. AI is the thing you're defending against Here's the map I ended up drawing for myself on the second one: • Identity - agent identity, sub-agent identity, identity lifecycle, provenance and lineage, delegation • Access - authentication, authorization, credential management, agent-to-agent trust • Governance - multi-agent orchestration governance, policy-as-code, human approval and escalation, compliance • Runtime - sandboxing, action interception, tool/MCP security, runtime policy enforcement, secret and data protection Accountability - immutable audit trail, action attribution, behaviour monitoring, incident response Every vendor I spoke to, had unclear answer on these 5 categories But what is a standard when you have 7 bodies building their own framework solving in accordance to solve core problem: • OWASP has a Top 10 for Agentic Applications • CSA has an Agentic Trust Framework • NIST launched an AI Agent Standards Initiative in February • OpenID Foundation has a draft for agent identity claims • IETF has a draft treating agents as workloads • Linux Foundation picked up SAFE-MCP • CISA and Five Eyes put out guidance in May Which means every vendor on that floor can point at a standard and be telling the truth. They're just each pointing at a different one, covering a different slice of that map What happens in vegas, goes on Twitter I guess ;) #BlackHat2026 #DEFCON #AISecurity #SecurityforAI – at Paradise, NV

I created this open source version of Clippy for mac with computer use and more But it is well guarded by @prismor_dev , so shouldn't do any cybercrimes github.com/Ar9av/clippy

The Hugging Face x OpenAI incident was exactly what was predicted decades ago Worth reading this paper from 1960 "Some Moral and Technical Consequences of Automation" where Norbert Wiener warned that machines can develop strategies that "baffle their programmers" cs.umd.edu/users/gasarch/BLO…

We internally benchmarked @prismor_dev classifier with deterministic scoped rules against auto mode of Claude Prismor outperforms across: • Goal hijacking • Tool misuse • Code execution • Memory poisoning Classifier + enforcement > classifier alone go use it for free

Prismor: Open Source Enterprise Control Plane for AI agents 🛡️ Running Claude Code, LangChain or BrowserUse Agents? Prismor checks every tool call against policy before it executes. 1) Blocks destructive commands and secret exfiltration inline. You pick enforce or observe per device. 2) HITL approvals: on a step-up rule the agent waits, a teammate taps Approve in Slack, the agent continues. Wire the same event to any webhook and build your own approval bot 3) guard_tools([...]) covers your whole toolset. 12+ frameworks, plus hooks for Claude Code, Copilot, and Codex. 4) MCP gateway: one connector governs every MCP server you use. 5) Tamper evident : Ed25519-signed audit trail ⭐ github.com/PrismorSec/prismo…

Fable is out AGAINN and while it is the most capable model. We are trying to make sure your coding agents are much more secure we have prismor all opensource, would love to have all of you guys try this and give feedback: github.com/PrismorSec/prismo…