The AI Agent Finished the Job. It Also Followed the Attack.
PI-005 produced unauthorized tool calls followed by valid extraction results. A useful final answer can hide an agent going outside its job.
- AI Security
- Local AI
TECHNICALLY·ACCEPTABLE
Independent writing and experiments on AI agent security, private AI, and the architecture underneath them.
Featured Essay
All EssaysLatest
-
PI-005: The Model Wasn't Resisting the Attack
PI-005 now gives us much stronger evidence that the worker itself really was susceptible. When given a safe, contained action channel, the poisoned worker tried to use it in all 20 runs.
- AI Security
-
The Prompt Injection Worked. The Architecture Didn't Let It Matter.
The model stayed prompt-injectable. Splitting hostile input from privilege changed the result.
- AI Security
- Local AI
-
Scrapping the Setup and Trying Again
PI-003 ended inconclusive after the troubleshooting apparatus became its own failure. I am replacing the runtime, qualifying it once, and rerunning PI-002 as a new experiment.
- AI Security
- Local AI
-
Troubleshooting Model Timeout
PI-003 localized Sentinel’s completion problem to model generation after document retrieval. The protocol still says the result is inconclusive.
- AI Security
- Local AI
-
A Timeout Is Not a Security Result
A failed run is not a blocked attack, and poisoned input can damage an agent without exfiltrating data.
- AI Security
- Local AI
-
Migrating from Claude to Codex: Claude considers me a security risk.
Why I moved from Claude to Codex after Claude's safeguards repeatedly blocked ordinary AI-security lab planning and writing.
- Tooling
-
Building a Prompt-Injection Test I Can Actually Trust
Sentinel now has matched corpora, a contained egress path, and a cleaner experimental boundary.
- AI Security
- Local AI
-
The Agent Runs Now
What runs, what does not, and why one apparent refusal still proves nothing.
- AI Security
- Local AI
Programs
Sustained technical work that produces the experiments, systems, and evidence behind the writing.
AI agent security
Sentinel
Controlled experiments in prompt injection, privileged tool use, architectural containment, and agent reliability.
Current questionDoes poisoned input increase accepted unauthorized send_email attempts when a worker is given a contained action channel?
Explore SentinelField Guides
Durable technical reference material for building, operating, and securing systems.
01 →
CMMC
A plain-English guide to scoping CMMC and meeting the level your contracts require.
In Progress GRC / CMMC
02 →
Cyber Security
Practical security work: labs, write-ups, and defenses that hold up.
In Progress SECURITY
03 →
Home Lab
Practical patterns for network segmentation, service isolation, and self-hosted infrastructure.
In Progress NET / INFRA
04 →
AI & Automation
Building AI workflows and agents, plus the automation and tooling that powers real work.
In Progress AI / AUTOMATION
About
Travis Shea
I lead complex technology programs and independently research AI agent security, private AI, and technical system design. Technically Acceptable is where I publish the resulting experiments and judgment.
About Travis