Home / Docs / Security model
Security advisories
Every security fix of a published Una release that affected an earlier published release, newest first. Each entry says what an accepted program could do, which releases were affected, which release of each line fixed it, and where the fix is described. The source of every entry is the release's section of CHANGELOG.md and the specification and design decisions it cites; this page paraphrases them and links them. To report a vulnerability, see SECURITY.md.
A fix made before any release that it would affect is not an advisory. Neither is hardening that no accepted program reaches. Both are in CHANGELOG.md, and the cases this page considered and left out are listed at the end.
Severity
- Critical: an accepted program can move
secretdata, or a decision on it, to a public sink through an ordinary construct. - High: an accepted program breaks a stated guarantee, noninterference or authority, in a narrower shape, or the run uses a handler the checker did not check.
- Medium: an accepted program reveals
internal-labeled facts, or one bit decided by them. - Low: details of the machine or account that built a release are exposed in its artifacts.
- Availability: a denial of service: an accepted program that reads untrusted input can be made to use resources, time or memory, without bound or superlinearly in the size of that input.
"Affected" names the releases that CHANGELOG.md names. Where it names none, the entry says so, rather than infer a range.
Advisories
UNA-SA-019: the v0.1 proof cache accepted entries it had not written
- Severity: High
- Affected: every published release of the v0.1 line up to and including v0.1.12, at tier 1 (
una check, the build report), in edition0.1. Releases from v0.2.0-alpha.3 on, including the v0.3 line, authenticate their cache (ADR 0064) and are not affected. - Fixed in: v0.1.13.
- Summary: the v0.1 verifier kept the solver's answers in
.una/cache/proofs/inside the package and reused any entry whose name matched a query. The name is a hash ofverification.lockand of the query, both derived from the package itself, and the entries carried no authentication. A repository could therefore ship cache entries marking any proof obligation as proven, a false one included. Runninguna checkin a clone of it then reported those contracts as proved without running the solver, and the build report counted them the same way. Entries now carry an HMAC-SHA-256 tag under a key that belongs to the user and lives outside the package (the fileUNA_PROOF_KEYnames, else~/.una/proof-cache.key, made on first use, readable by the user only). An entry without a valid tag is ignored and its query solved again, so caches written by earlier releases are rebuilt once. The build report names the key file instead of listing the cache as trusted. The key file is used only when it is a regular file with one link, owned by the user, granting nothing to group or others, of 16 to 4096 bytes; otherwise the cache is off. - References: #1164, ADR 0064.
UNA-SA-018: una fmt followed a symbolic link planted at its temporary file
- Severity: High
- Affected: every published release up to and including v0.3.16 and v0.1.12:
una fmtin both lines, anduna fixin the v0.3 line, in every edition. - Fixed in: v0.3.17; v0.1.13.
- Summary:
una fmtreplaced a source file by writing a temporary file at a fixed name next to it (.<name>.fmt-tmp) and renaming it over the file. The write followed a symbolic link at that name. A repository can commit such a link, so runninguna fmtin a clone of it could overwrite any file the user can write, outside the repository, with text the repository chose, and leave the source file as a link to it.una fixin the v0.3 line used the same write, and the proof cache thatuna checkkeeps in.una/cache/proofs/wrote its entries the same way. The temporary file is now created exclusively, under a name unique to the write, so nothing already at that name is written through, and the proof cache stores nothing when.una,.una/cacheor.una/cache/proofsis a link that leads outside the package. A source file that is itself a symbolic link is still formatted in place only when the file it resolves to is inside the package; a link leading outside it is refused. Users of an affected release should not rununa fmtoruna fixin a repository they do not trust until they update. - References: #1160, #1161, #1162, #1163.
UNA-SA-017: releasing a deeply nested value overflowed the interpreter's stack
- Severity: Availability
- Affected: every published release up to and including v0.3.16 and v0.1.12, in the interpreter (
una run), in every edition. Native executables are not affected. - Fixed in: v0.3.17; v0.1.13.
- Summary: the interpreter released a value that a function value closes over by recursion on the native stack, one level per level of nesting. A loop that nests values through function values, such as a kind
L(next: fn() -> F)built two million levels deep, finished its work and then aborteduna runwith a stack overflow when the value was released, instead of ending normally. A program whose nesting depth followed untrusted input could be made to abort this way. Function values now release the values they close over and the handlers of their capabilities by the same worklist that releases records. The v0.1 line had no worklist at all, so its fix gives one to records and function values together. There, reaching the overflow at the default--max-depthtook a value of about 650,000 levels, which costs quadratic time to build, and a lower--max-depthreached it sooner. - References: #1071, #1076, ADR 0060 rule R5.
UNA-SA-016: parsing a long integer from text took time quadratic in its digits
- Severity: Availability
- Affected: every published release up to and including v0.3.15 and v0.1.11:
to_intin every edition, andparse_intfrom edition 0.2. - Fixed in: v0.3.16; v0.1.12.
- Summary:
to_intandparse_intconverted a text of d digits to anIntin O(d²) time, so a program that parsed untrusted text could be made to spend quadratic time in one call. Both now refuse a text of more than 4300 digits, the sign excluded and leading zeros included, after one O(b) scan and before any conversion:to_intreturnsNone, andparse_intfails with the newParseErrorvariantTooLong, the bound that JSON integers already had. AwhenoverParseErrorthat names each variant without anelsearm must now handleTooLong. The v0.1 line has noparse_int, so its fix boundsto_intalone. - References: #914, #922, #1027, spec/22 §4, spec/14 §1, spec/23 §3.
UNA-SA-015: an interpolated text pattern let a secret choose a branch or a public Bool
- Severity: High
- Affected: every published release up to and including v0.3.14 and v0.1.10, in every edition.
- Fixed in: v0.3.15; v0.1.11.
- Summary: a
whenarmis "…"and the expressionx is "…"treated their text literal as apublicconstant, but a literal with interpolation is computed from the values it interpolates. Withs: secret Text, awhenover another value with an armis "{s}"could choose the arm that prints, andx is "{s}"gave apublicBool, with no diagnostic. Interpolation in a pattern's literal is now U4008 in every edition; comparing with==gives aBoolthat carries the value's label. - References: #801, #803, spec/07 §7 rule 3.
UNA-SA-014: release archives named the account that built them
- Severity: Low
- Affected: every published release archive through v0.3.13 and v0.1.10.
- Fixed in: v0.3.14. The v0.1 line is not fixed.
- Summary: each entry of a release archive carried the user and group names of the account that built it. The binaries in the archives were not affected. Archive entries are now owned by
0:0with no user or group name, andscripts/check-release.shchecks the owners. - References: #679, #689.
UNA-SA-013: a function named as a value ran on the bindings where it was named, unchecked there
- Severity: High
- Affected: every release before v0.3.13 and v0.1.10.
- Fixed in: v0.3.13; v0.1.10.
- Summary: the checker took a named function value's
usesonly from the code that called it. A reference with no binding of a name its function uses was accepted and stopped the run, and a reference under one handler that was called under another was checked against the second and ran on the first. A reference now needs the function'suseswhere it is written: U3001, or U3003 whereusesare declared empty. - References: #565, #673, S-20, ADR 0093 amendment of 2026-10-14, spec/05 §6 rule 7, spec/11 §4.
UNA-SA-012: a policy call's failure was public whatever the policy's to label
- Severity: High
- Affected: releases before v0.3.7 and v0.1.9;
CHANGELOG.mdnames no first affected release. - Fixed in: v0.3.7; v0.1.9.
- Summary: a
viafunction that failed on itsfromvalue let the caller see the failure as public, so a public branch onattempt p(v)revealed what it decided about a secret. The failure is now labeledpc ⊔ ℓ(to). - References: #496, S-18, ADR 0090.
UNA-SA-011: a checked predicate's failure did not carry the labels of what it read
- Severity: High
- Affected: releases before v0.3.6 and v0.1.8;
CHANGELOG.mdnames no first affected release. - Fixed in: v0.3.6; v0.1.8.
- Summary:
checked(a, φ) else fail Clabeled its failure byaalone, so a predicate that read a secret, asraw.len < limitwith asecretlimit, failed publicly. The failure now carries every label that φ reads. - References: #488, S-17, ADR 0089.
UNA-SA-010: whether host text was valid UTF-8 could reach a public branch
- Severity: Medium
- Affected: releases before v0.3.5 and v0.1.7;
CHANGELOG.mdnames no first affected release. - Fixed in: v0.3.5; v0.1.7.
- Summary:
Files.readfailed andEnv.getreturnedNoneon text that was not valid UTF-8, outcomes decided byinternalcontent that a public branch could see. Both now abort the run with U6005. - References: #460, S-16, ADR 0086, spec/14 §3.
UNA-SA-009: an HTTP request's failure could reveal internal data the server received
- Severity: Medium
- Affected: releases before v0.3.4, in edition 0.2;
CHANGELOG.mdnames no first affected release. The v0.1 line has noHttp. - Fixed in: v0.3.4.
- Summary: a server could decide
TooLarge,ProtocolorTimedOutby a request's body or header values, but the failure was labeled by the URL, the header names and the key name only. Every argument now labels it, so a request's failure isinternal. - References: #437, S-15, ADR 0085, spec/24 §5.
UNA-SA-008: a write to a labeled path leaked through later file failures
- Severity: High
- Affected: releases before v0.3.4 and v0.1.6;
CHANGELOG.mdnames no first affected release. - Fixed in: v0.3.4; v0.1.6.
- Summary: whether a later read succeeded revealed whether a secret path had been written, in the same run or a later one. Every
Filesfailure now also carries the package's file-state label. - References: #436, S-14, ADR 0085.
UNA-SA-007: executables and release archives named paths of the machine that built them
- Severity: Low
- Affected: releases before v0.3.3;
CHANGELOG.mdnames no first affected release, and does not name the v0.1 line. - Fixed in: v0.3.3.
- Summary: on macOS,
una buildkept the absolute path of the runtime library, usually with the home directory and the user name, in every executable, and release archives named paths of the machine that built them. Executables now link without debug information, and releases remap those paths. - References: #382, S-13.
UNA-SA-006: in edition 0.1, an annotation dropped a decision on T values
- Severity: Critical
- Affected: v0.1.0 to v0.1.4; edition 0.1 packages checked by releases of the v0.3 line before v0.3.2, whose first affected release
CHANGELOG.mddoes not name. - Fixed in: v0.3.2; v0.1.5.
- Summary: in a generic body,
let y: T = if a == b then p else q, or a fieldHold[T], dropped the label of the comparison, which a later call could reveal. A decided value flowing into an annotation, a field or a function value's parameter of typeTis now U4001. - References: #354, #345.
UNA-SA-005: in edition 0.1, a decision on T values given to function values
- Severity: Critical
- Affected: v0.3.0, for edition 0.1 packages; v0.1.0 to v0.1.4.
- Fixed in: v0.3.1; v0.1.5.
- Summary: a call that gave function values data decided by comparing values of
Tcould reveal the decision, through the prelude's higher-order functions or a generic function of the program. Such a call now runs under the decision and raises its result. - References: #346, #350.
UNA-SA-004: in edition 0.1, a decision on T values through a call of a function value
- Severity: Critical
- Affected: v0.1.0 to v0.1.3.
- Fixed in: v0.3.0; v0.1.4.
- Summary: a value chosen by comparing values of
T, passed to a function value whose result reached a sink, was accepted. It now raises the result and the program counter of the call. - References: #253, #255.
UNA-SA-003: whether Files.write failed revealed an internal path
- Severity: Critical
- Affected: v0.1.0 to v0.1.3; whether the v0.2.0 pre-releases were affected,
CHANGELOG.mddoes not say. - Fixed in: v0.3.0; v0.1.4.
- Summary: a host operation's failure is now labeled by the arguments that decide it, so
attempton a write to aninternalpath gives aninternalResult. - References: #310.
UNA-SA-002: in edition 0.2, a decision on T values given to an effectful uses of parameter
- Severity: Critical
- Affected: v0.2.0-alpha.3; whether earlier pre-releases were affected,
CHANGELOG.mddoes not say. - Fixed in: v0.3.0.
- Summary: a value chosen by comparing values of
T, passed to auses ofparameter whose effect revealed it, was accepted. Each caller now runs its argument under the labels of the type argument. - References: #234, ADR 0067.
UNA-SA-001: in edition 0.2, function values given for of parameters lost their labels
- Severity: Critical
- Affected: v0.2.0-alpha.3. The v0.1 line is not affected.
- Fixed in: v0.3.0.
- Summary: a function value given for a
uses oforfails ofparameter, asmap,filter,first_whereandfoldtake, lost its labels: a leak, and a false rejection. - References: #210.
Not advisories
- #1514 (S-41), #1515 (S-42) and #1544, the trace of a run showing how many handler installations and operations ran under a branch on data labeled above
public(v0.4.0): in every published release of the v0.1 and v0.3 lines and in the v0.2 alphas, in every edition they serve, this is the specified behavior. Each operation invocation appends one event, and each handler installed bywithhas the idwith:<handler>#<n>(spec/11 §5); the count behind<n>runs across the run, by spec/19 §7 rule 2 from release 0.2.0-alpha.3 on, and by the implementation in the v0.1 line and in releases 0.2.0-alpha.1 and 0.2.0-alpha.2. Noninterference (spec/07 §1) is stated per observer label. An operation of the program's effects under such a branch, directly or through a call, needspc ⊑ accepts(spec/07 §3 rule 4), so whoever sees its event already sees the branch. Awithdischarges the effects of the capability its handler serves (T-Handler, spec/10 §6), so an operation that handler serves is no effect of the program; from release 0.3.8 on, no observer of the formal model sees it (F1EvVis). No released document or proof states that instance ids,seqor the number of events are independent of labeled data. "Traces never show labeled data", in ADR 0032 and in the v0.1 line'sREADME.md, is about the values a trace renders (ADR 0032 rule 3), and the sameREADME.mdsentence says that every effect is recorded in the trace: these findings render no labeled value. An argument whose static label is abovepublicis redacted (ADR 0032 rule 3), and what varies with the branch is which events occurred, how many, how they are numbered, and whichpublicvalues their arguments carry. Outcomes that the trace wrote in clear are a separate exposure, S-45 (#1552). v0.4.0 makes it a rule, applied in every edition, that quiet code numbers no installation and appends no such event (spec/26 §6 rule 4). - #773 and #775, native code ending with an allocator abort where the interpreter reports U6003, for a wide range in a
together forand for a copy between tasks (v0.3.15): memory stayed bounded and no label or authority guarantee was broken, so no severity of the scale applies. - #779, rustls 0.23.46 for four published advisories (v0.3.15): none is reachable in the TLS configuration of
Http. - #799, the labels of some values in the Core export (v0.3.15): the export is read by the formal re-check, never by a run; no accepted program leaked through it.
- #323, the labels of an un-annotated
let edit(ADR 0075):CHANGELOG.mdnames only v0.2.0 as affected, which was not published. - #310's
Httppart, and HTTP credentials kept to their origin (v0.3.0):CHANGELOG.mdnames no published release that hadHttpbefore the fix. - #186,
editarguments flowing back with their final labels:CHANGELOG.mdnames no affected published release. - #216, defense in depth near
ofparameters: none of its sites leaked. - #424, the runtime checks a task's code before calling it, and #480, native code aborts on a
setoutside the list: hardening that no accepted program reaches.