What is open, how fast are we responding, and where should we look next? · For Security
7 findings are open, 5 of them critical or high. 9 were new in this window and 21 were resolved, 5 days to resolve on average (17 resolved). 3 were blocked and 6 only observed.
- New
- 9
- Resolved
- 21
- Avg. to resolve
- 5 d
- Blocked
- 3
Open now
as of Nov 1What is open right now, and what deserves attention?
Triage severe or aging Findings.
| Severity | Finding | Policy | Team | Age | Outcome |
|---|---|---|---|---|---|
| critical | AWS access key AKIA…7Q2F in a prompt at 11:47, from a pasted .env file; sent to the model | Credentials must not be sent to a model | Engineering | 25 d | Observed |
| high | A social security number reached the model unsanitised | Sensitive information sanitized before transmission | Engineering | 42 d | Observed |
| high | A card number reached the model unsanitised | Sensitive information sanitized before transmission | Finance | 42 d | Observed |
| high | A bank account number reached the model unsanitised | Sensitive information sanitized before transmission | Team B | 41 d | Observed |
| high | Provider account usera@gmail.com, not organisation-backed | Personal account is in use | Team A | 24 d | Observed |
| medium | github.merge_pull_request ran without approval: #517 into main, 14 files | Consequential tool calls require approval | Engineering | 27 d | Observed |
| medium | Instruction override in issue #482 "Build fails on arm64", read through GitHub MCP (score 0.91) | Prompts are inspected for injection and leakage | Engineering | 26 d | Observed |
The 7 highest-priority of 7 open, by severity, then age.
/api/security · findings · always as of the data cutoff, whatever the period
Resolved
in windowstagedWhat was resolved during the period?
See what work was completed and why Findings closed. Resolving isn't always a technical fix.
- Resolved
- 21
- By a person
- 1
- By a policy change
- 20
- Average to resolve
- 5 d
- 17 findings
- 15 now enforced
- 3 prevented
- 2 allowed
- 1 By design
/api/security?period · resolutions
Backlog
over the windownot in qdash yetIs the backlog growing or shrinking?
See whether investigation capacity is keeping up.
9 new, 21 resolved, 3 reopened. Backlog 16 at the start, 7 at the end.
finding status events · needs status history
Time to resolve
over the windownot in qdash yetAre we responding faster?
Spot changes in response time without hiding the unresolved backlog.
From first detection to the latest resolve in the period, for each finding resolved at the data cutoff. A finding reopened and still open is backlog, so it isn't counted here.
finding status events
Most violated policies
in windowWhich policies are violated most?
Prioritize investigation or policy review.
| Policy | Findings | Worst | Teams |
|---|---|---|---|
| Unsanctioned MCP server in use | 3 | high | 3 |
| Personal account is in use | 1 | high | 1 |
| Credentials must not be sent to a model | 1 | critical | 1 |
| Prompts are inspected for injection and leakage | 1 | medium | 1 |
| Consequential tool calls require approval | 1 | medium | 1 |
| Restricted file access | 1 | high | 1 |
| Approved model providers only | 1 | medium | 1 |
Findings first detected in the window. Top 7 of 7 policies.
/api/security?period · by policy
Blocked and observed
in windowWhat was blocked versus only observed?
See recorded blocking and observed exposure. A block isn't the same as a resolved Finding.
| Severity | Enforced | Observed |
|---|---|---|
| critical | 0 | 1 |
| high | 2 | 3 |
| medium | 1 | 2 |
| low | 0 | 0 |
| All | 3 | 6 |
Counts Findings first detected in the window, not attempts or unique attacks. The outcome is what happened at detection: later policy changes don't rewrite it.
finding outcome at detection
Corrections
as of Nov 1stagedWhich findings were answered by a completed training?
See where a finding became a correction. Names stay on the finding and in the audit log, not here (c133 D10).
- Assigned
- 8
- findings with a training
- Completed
- 0
- 0 in this window
- Waiting
- 8
- nothing is blocked meanwhile
- Unapproved MCP serversT-MCP-10 of 3 completed
- Personal accounts for workT-ACCT-10 of 1 completed
- Secrets in promptsT-SECRETS-10 of 1 completed
- Instructions hidden in contentT-INJECT-10 of 1 completed
- Ask before an agent sends or publishesT-APPROVAL-10 of 1 completed
- Approved model providersT-PROVIDERS-10 of 1 completed
Who completed what is on each finding and in the audit log, not here.
training.completed.* · training-assignment messages (c133, proposed) · the lab counts this visit's completions
Data protection
in windowWhich sensitive paths and secrets did agents touch?
See where sensitive data is exposed to agents.
- Credentials and keys214
- Source and config182
- Customer data132
- HR and finance101
/api/dlp?period · seeded in the lab
Egress decisions
in windowWhat left the fleet, and what was stopped?
See risky destinations, and what enforcement stopped.
268 risky requests; 201 blocked (75%), 67 let through where policies only observe.
Where the risky traffic went- pastebin.compublic paste83
- files.catbox.moeanonymous upload64
- chat.openai.com (personal)personal account62
- 203.0.113.42unknown host59
/api/egress-decisions · trends egress · seeded in the lab
Enforce-pilot candidates
in windowstagedWhere could we start enforcing?
Pick a manageable pilot. Nothing here changes a policy.
- Agents may not run as root
Consider enforcing this for Engineering. Over these 30 days, Qpoint monitored 1,608 sessions with no observed violations. Start with a pilot.
Engineering · 100% of its endpoints covered · 1,608 monitored sessions · 0 observed · 0 exceptions - Tools must be on the allow list
Consider enforcing this for Engineering. Over these 30 days, Qpoint monitored 1,403 sessions with no observed violations. Start with a pilot.
Engineering · 100% of its endpoints covered · 1,403 monitored sessions · 0 observed · 0 exceptions - Agents run in a sandbox
Consider enforcing this for Engineering. Over these 30 days, Qpoint monitored 1,403 sessions with no observed violations. Start with a pilot.
Engineering · 100% of its endpoints covered · 1,403 monitored sessions · 0 observed · 0 exceptions
Few findings alone don't show a rollout will be painless, so start small. A recommendation never turns a policy on.
rule over mode, findings, sessions and coverage · monitored sessions are seeded