AI Safety + Cybersec
Gonzalo Del Castillo
Cybersecurity -> AI Safety
Gonzalo Del Castillo
I am building a path into AI safety from the intersection of cybersecurity, tool-using agents, evaluations, and reward hacking.
Selected Direction
A matched-pair measurement of what removing an LLM's refusals actually does to its offensive capability. Short version: stripping refusals does not make a small local model a more capable attacker, and the effect flips sign depending on whether the task is bound by capability or by willingness.
Read the summary →Small environments where an agent can optimize the wrong metric, manipulate the evaluator, or exploit tool access to "win" without solving the real objective.
Applying a pentest mindset to autonomous agents: memory abuse, tool misuse, permission escalation, and broken incentives.
A focused archive of cybersecurity, networking, defense, and fundamentals certificates. Open the vault to inspect each one.
Open certificatesA biometric voice pipeline that checks both speaker identity and the spoken password, evaluated with FAR/FRR.
Open project summaryBuilt with Python and PyTorch over 45k+ WAV files, 816 speakers, and 5 spoken passwords. The project trains a speaker embedding model and a phrase classifier, then evaluates four verification scenarios: genuine access, wrong phrase, impostor with correct phrase, and full impostor.
View GitHub repositoryAbout
My background is not a straight research-lab pipeline, and I am trying to use that honestly. Cybersecurity has trained me to look for incentives, boundaries, abuse paths, and failure modes before systems are trusted too much.
Right now I am focused on autonomous agents, tool-use, evaluations, reward hacking, and small research projects that are concrete enough to test and publish.
Skill Tree
Your base route: AI safety with a cyber mindset.
Offensive fundamentals and attack surfaces.
Mapping abuse, permissions, and trust boundaries.
Networks, segmentation, and defensive basics.
Attacker thinking, practiced carefully.
Industrial security, operational risk, and control.
Agents with tools, memory, and actions.
Containing files, tools, and execution.
When the metric separates from the real goal.
Measuring behavior, not just answers.
Capstone: agents, evals, oversight, and control.
Small, public, and traceable experiments.
Monitoring and limiting agentic systems.
Minimum permissions for tools.
Local community, workshops, and public signal.
Notes
A reflection on emotional and cognitive intelligence in humans, and what emerges when the decision-maker is a machine.
Read the essay →Contact
