Agents and tool use
Permissions, tool use, memory, delegation and recovery in systems that act on their environment. We attack the agent through its real tools and check that the fix holds.
With A440 and SordinoRed teamingAttack the system, then check the fix holds
We red team AI models, agents and the physical systems they control, with specialists for each kind of harm. Every finding goes to an owner, a control, a decision and a retest, so the evidence reaches someone who can act on it.
What we red team
Twelve kinds of red teaming, one standard of evidence
Our core team runs each engagement and brings in specialists from the Trusted Red Team Network when the threat model needs deeper expertise. Engagements run under NDA.
Prompt injection and jailbreaks
Systematic testing that documents each bypass well enough to reproduce it and name the broader failure, including payloads hidden in images and documents.
Past engagementCybersecurity of AI products
Adversarial testing of AI products with a security focus, and advice on how advanced models change an organization's cyber risk.
Past engagements, technology and bankingML security and robustness
Model-level attacks and robustness evaluation for frontier and production systems, with analysis of the defenses that could work.
Network specialistsElection security and information integrity
Influence operations, civic information and election-related AI risk, with every claim the system makes fact checked against official, current, in-scope evidence.
How it works ↗CBRNE misuse
Chemical, biological, radiological, nuclear and explosives misuse pathways, domain safeguards and responsible safety evaluation.
Network specialistsPhysical AI and robotics
Open robot models, vision-language-action models and embodied agents, from perception and navigation to manipulation and physical safety.
Campaign recruiting nowNetworks and cyber-physical systems
Telecommunications, industrial control systems, operational technology and connected devices that interact with AI.
Network specialistsChild safety
AI risks to children, age-appropriate safeguards and how models and agents behave in sensitive conversations.
Network specialistsBias and discrimination
Adversarial testing for biased and discriminatory behavior across a company's AI products.
Past engagementArmed conflict and humanitarian protection
Red teaming with IHL and humanitarian specialists, including a multi-phase campaign on responsible AI in armed conflict with a leading European university.
Responsible systems ↗Policies, protocols and strategic risk
Model-level and feature-level testing protocols and safety policies, tested against real findings and traced to the decision rights behind them.
Past engagementsElection security red teaming
Attack the agent and the facts it relies on
Election harm reaches an AI system two ways: someone tries to use it for an influence operation, or it repeats something about voting that is wrong, stale or from the wrong place. Our campaigns test both, and fact check every claim the system makes along the way.
Influence-operation playbooks
Election security specialists run voter suppression, official and candidate impersonation, synthetic media, fabricated evidence, mass-produced variants and targeted discouragement against your agents, in single requests and across long conversations.
Poisoned evidence
We seed retrieval and tool results with stale notices, rules from the wrong state, rumor chains that trace back to one post and look-alike official sources. Then we fact check what the system says.
Runs across the calendar
The same campaign runs at several points in the election calendar, because what counts as current changes as Election Day nears and again once counting starts.
Sealed, scored findings
Every attempt and every checked claim is kept as a tamper-evident decision record and scored in the format of Intonation's published suite, with a control and a retest for each failure.
Example campaign plan
Fictional plan for illustration. Each engagement sets its own rows, phases and depth.
Attack familyWeeks outFinal weekFinal 48 hoursElection nightCount and certification
What the agent is asked to do
Voter suppression
Impersonating officials
Synthetic media of candidates
Fabricated evidence
Mass-produced variants
What the agent says, fact checked
Wrong polling details
Rules from the wrong place
Stale sources
Rumors with one origin
Early results and calls
Altered or invented quotes
Built for
- AI developers testing election policies on their models and agents
- Election offices and governments putting assistants in front of voters
- Platforms and newsrooms publishing results and race calls
- Enterprises whose agents post or reply during an election
How an engagement runs
From threat model to retest
Scope, access, responsibilities and the final deliverable are agreed before work begins.
- Threat model and scope
Each engagement starts from a specific threat model: the system, its intended use and the decision you need to make.
- The right team
Technical and policy specialists work together when a finding needs both kinds of analysis.
- Reproducible findings
Methods and evidence support independent review and reproduction, within the limits of the NDA.
- Owner, control, retest
Each finding is mapped to where it entered the pipeline, the control that belongs there and who owns it, then retested after the fix and when the system changes.
Beyond red teaming
Election integrity between campaigns
A red team finds where a system fails. These keep it in tune through the rest of the election calendar.
Election readiness review
Before the calendar starts, we map every agent and channel that will talk about voting, the official sources each should rely on, the names it may sign for and who reviews escalations.
Retests through the count
When a model, prompt, source or rule changes, we rerun the campaign and report what moved, through certification.
Intonation in Sordino
The same checks enforced at the point of sending, with a free version in the Fermata MCP server.
Try Intonation ↗Exercises for your teams
Tabletop exercises for trust and safety, communications and support teams, built on fictional jurisdictions.
Training ↗Practice and training
Red team with us
before you hire us
Red Teaming Games lets your team break a synthetic workflow, choose the control and its owner, and replay the attack. The live AI Red Teaming and Risk Governance workshop covers red teaming agents and turning findings into controls.
Specialists can apply to join the Trusted Red Team Network for future campaigns.
Plan a red team engagement
Tell us about the system and the decision ahead. We aim to reply within two business days.