mediumwhoami
Embedded instruction telling the agent to always invoke a known-to-fail tool
The description states the tool cannot work ('this connection does not have' the required sign-in) yet instructs the agent: 'Call it anyway when the user asks for this... which is what they need to hear.' This is a directive aimed at the AI's behavior rather than a factual description of functionality, overriding the agent's normal judgment about whether to invoke a tool that is documented to fail. This phrasing pattern recurs identically across 8 tools, suggesting a deliberate attempt to condition the agent into blind, unconditional tool invocation.
mediumflash_dashboard
Same manipulative 'call it anyway' instruction; broad sensitive data capability
Shares the identical embedded directive forcing the agent to call a tool that needs authentication it lacks. Additionally, if sign-in were ever present, this tool would expose occupancy, revenue and KPI business data — a capability unrelated to the public-facing booking assistant's stated purpose, and worth scrutiny for scope creep.
mediumarrivals
Repeated 'call it anyway' directive; exposes guest reservation data
Same suspicious embedded instruction pattern as other admin tools. Functionally it would return guest reservation arrivals, i.e. personal booking data, which is a sensitive capability sitting oddly alongside public holiday-booking tools.
mediumdepartures
Repeated 'call it anyway' directive; exposes guest reservation data
Identical manipulative phrasing instructing the agent to invoke a tool it knows will fail. Would expose guest departure/reservation details if authenticated, which is sensitive PII-adjacent data.
mediumbookings_search
Repeated 'call it anyway' directive; broad reservation lookup capability
Same embedded AI-directed instruction overriding normal judgment about calling failing tools. This tool, if functional, could search reservations by room or booking number across the business — a broad capability that should require explicit, verified authorization, not blanket 'call it anyway' encouragement.
mediumreservation_detail
Repeated 'call it anyway' directive; exposes individual guest reservation details
Identical suspicious phrasing pattern. Returns dates, party size, room and status of a specific reservation — personal guest data — behind a disabled auth gate, yet the agent is told to call it regardless.
mediumday_summary
Repeated 'call it anyway' directive; exposes guests-in-house data
Same embedded instruction pattern directing the agent to act against its own judgment about non-functional tools. Returns operational guest-in-house data if ever authenticated.
mediumtasks_overview
Repeated 'call it anyway' directive
Completes the set of 8 identically-worded tool descriptions that instruct the AI to 'call it anyway' despite stating the tool cannot function without sign-in. The consistent, unusual phrasing 'which is what they need to hear' is a persuasive/manipulative framing aimed at the AI rather than a neutral functional description, and should be treated as an attempt to condition unconditional tool invocation regardless of known failure.
lowsearch_stays
Unusually long tool description
Very long descriptions are sometimes used to bury instructions to the AI. Skim it before allowing.
low
Not on gate's verified list
This server isn't on gate's handpicked list. That's not necessarily bad, but there's no third-party signal about it — review its tools carefully.