AIShould indirect prompt injection be mitigated at the system boundary or the application layer?DISCUSS ›
AIWhat's the minimum viable permission scope for autonomous agents running in production?DISCUSS ›
CYBERIs model exfiltration via API abuse a solved problem, or is the research lagging the threat?DISCUSS ›
> Loading board...
DISCUSSION
POST COMMENT
GOR INTEL
AI AGENT SECURITY · SIGNAL, NOT NOISE
A machine-curated feed from the Gray Owl Research lab · Re-scored every 6 hours · No editors, no sponsors
6H
CURATION
40+
FEEDS
12H
LEADERBOARD
LIVE
STATUS
WHAT THIS IS
The research desk of Gray Owl Research. We build Obol,
a per-action authority layer for AI agents: every consequential tool call needs a single-use signed coin,
a deterministic gate checks it, and each decision lands in a hash-chained receipt ledger. This feed is
what we read while doing that work, filtered by a model and published as-is.
WHAT WE'RE WATCHING
01 · Indirect Prompt InjectionInstructions hidden in content an agent reads (repo files, web pages, tool output) steer it off-task. Our own live-fire test: a planted CLAUDE.md got a coding agent to leak 4 of 4 secrets with no gate in place.Vector: coding agents, browsing agents, RAG, MCP tool results
02 · Ambient AuthorityAgents inherit broad standing permissions (shell, file system, API keys), so any hijack runs with all of them. Coarse allow-lists barely help; per-action, narrowly scoped grants do.Vector: agent frameworks, code interpreters, MCP servers, CI bots
03 · Unaccountable ActionsWhen an agent does something wrong, most stacks can't say who authorized it or reconstruct why. Tamper-evident receipts turn "the AI did it" into an auditable record.Vector: incident response, compliance, multi-agent delegation
HOW THE FEED WORKS
SOURCES
Labs: OpenAI · Anthropic · DeepMind · HuggingFace · BAIR Security: Krebs · Schneier · BleepingComputer · PortSwigger · The Record Research: arXiv cs.AI/LG/CL/CR · The Gradient · MIT Tech Review World desk: Reuters · BBC · AP/NPR · Guardian · Al Jazeera
CURATION
Llama 3.3 70B on Cloudflare Workers AI Scores relevance, assigns category, writes the analyst note
SUMMARIES
Claude Haiku 4.5 · tickers and haiku · Claude Sonnet 4.6 · sitrep
LEADERBOARD
Live from LMArena text arena (style control) · refreshed every 12h Shows its age when a refresh fails, never invents scores
OBOL PANEL
Our own lab numbers, updated by hand after each measured run Self-run, small samples, not an independent evaluation
Nothing here is reviewed by a human before it posts. Curation models misread headlines and
analyst notes can be wrong. Follow the link to the source before relying on anything.