Collection
| Limitation | Consequence | Practical response |
|---|---|---|
Only openat and connect activity is captured | open, openat2, sendto, sendmsg, accept, writes, renames, and unlinks are invisible | Do not infer absence from a missing event |
| Tracepoints are system-wide | Kernel produces host activity even in target-tree mode | Check filtered_events; treat scope as userspace admission |
| Ring buffer and userspace channel are bounded | Busy hosts can lose events | Check ringbuf_lost_events and userspace_dropped_events |
| Missing syscall enter/exit or LRU eviction | No activity event is emitted and it is not separately counted | Absence of an activity row is not proof of no attempt |
| Detached descendants may outlive the target | Collection may stop before their later activity | Keep the target alive or use an explicit waiting wrapper |
| Fixed tracepoint structs replace CO-RE | Kernel layout changes can require code and schema updates | Run scripts/check-btf-layout.sh before recording |
| BTF must be readable for layout verification | Some hardened or minimal kernels cannot be checked | Do not load if the layout cannot be verified |
| File paths are capped at 256 bytes | Long paths are truncated and marked | Treat a truncated path as incomplete evidence |
Identity and attribution
| Limitation | Consequence | Practical response |
|---|---|---|
| BPF and collector may see different PID namespaces | Raw IDs may not match observer IDs | Use canonical process identity and check unresolved flags |
Fork handler currently emits child ns_tgid=0 | Canonical child identity may wait for a later exec or exit | Expect provisional forks; do not invent ancestry |
/proc reads race with process exit | Start time or metadata may be unavailable | Check for fallback identity markers |
| Thread discrimination can race | A child may disappear before leader status is checked | Treat ambiguous identity as unresolved |
| PID reuse is resolved by start time | Two processes with the same PID remain separate instances | Use process-instance IDs, not PIDs alone |
| Parent linkage may be ambiguous | Tree edges can be missing or unresolved | Use tree notes and analysis limitations |
| Tracked process table is capped at 65,536 | Excess processes are filtered and counted | Watch filtered counts on fork-heavy workloads |
| Target-tree scope is not a security boundary | A hostile process cannot inject events but may influence attribution races | Do not treat scope as isolation |
Evidence semantics
| Limitation | Consequence | Practical response |
|---|---|---|
| An open is not a read | A file_open row proves only the attempt and result | Do not claim file contents were accessed |
| A connect is not a session | Success means connect returned 0 | Never report an established connection or data transfer |
| TCP and UDP are indistinguishable at this hook | Transport protocol is unknown | State that the protocol is not established |
| Source address is often unavailable | Source fields may be null | Do not infer the local endpoint |
| Relative paths are not resolved | Target file may differ from the string recorded | Use dirfd and context, and mark the limit |
| Timestamps derive from a captured offset | NTP steps can skew wall time | Prefer monotonic timestamps for ordering |
| Absence of an event is not proof of no activity | Loss and coverage gaps exist | Pair negative claims with collection accounting |
| No file contents, payloads, DNS, or environment variables | Some investigative questions cannot be answered | Use host-level tooling for those questions |
Integrity and custody
| Limitation | Consequence | Practical response |
|---|---|---|
| No tamper-evident log or signature | A writer can modify the database | Hash and custody records are external controls |
| Verification is not authenticity | valid means no observed contradictions | Do not equate validity with truth |
| Export manifest is not a signature | It detects post-export changes only | Record who handled the package |
| Analysis is not a verdict | Findings are deterministic indicators | Evaluate context, false positives, and limitations |
| No audit log of recorder operators | The tool does not record who ran sudo | Use OS audit tooling |
Operations
| Limitation | Consequence | Practical response |
|---|---|---|
| Recording requires privilege | Unprivileged users cannot capture | Use sudo or scoped capabilities |
| Root shell skips privilege drop | Target runs as root | Use sudo from an unprivileged account |
--redact-args covers arguments only | Paths and destinations remain | Treat evidence as sensitive |
| Read-only open does not migrate | Older or newer schema may fail | Use a compatible database; record a new session |
| Workbench has no record control | Recording is CLI-only | Record first, then open the database |
| Workbench export path is fixed | No save dialog or frontend path | Use SYSCALL_EXPORT_DIR at startup if needed |
| No signed installer or packaged bundle | Distribution is source-built | Build from a trusted checkout |
| Large desktop sessions are paged, not fully virtualized | Some large views may be slower | Filter by time, process, or search |
| No automated GUI end-to-end test | Visual regressions need manual checks | Run renderer and service tests plus a manual smoke test |
| CI cannot prove privileged kernel capture | Green CI is not a live-capture PASS | Run the operator e2e on a capable host |
What “no findings” means
A clean analysis or avalid verification does not prove that nothing happened.
- Collection may have missed activity.
- The relevant syscall may not be hooked.
- The session may cover the wrong process tree or time range.
- The rule set may not model the behavior.
- A path or identity may be unresolved.