Skip to main content

Glossary: AI coding agent failures and benchmark terms

Short definitions of the failure modes on this site, the other names they go by and the terms used to score detectors. Each entry has its own link; failure modes point to their full pages on the AI coding agent failure modes hub.

Failure modes

The modes SAICA-Bench defines, each linked to its page with measured results.

Cascading failure
A cascading failure is when an AI coding agent’s attempt to fix an error introduces new errors, and each further attempt moves the code further from a working state. A simple failure can become an irrecoverable tangle after repeated recovery attempts. Not benchmarked yet. Cascading failure: results, incidents and tools →
Context pollution
Context pollution is when an AI coding agent’s working context degrades over a long session: relevant earlier information is dropped, irrelevant material piles up, or mistakes from earlier turns are treated as fact in later ones. It is a failure of state hygiene across a session, not of a single generation. Not benchmarked yet. Context pollution: results, incidents and tools →
Dependency blindness (reinventing existing code)
Dependency blindness is when an AI coding agent writes its own version of a function, class or utility that the standard library, an installed package or the codebase already provides, for example a hand-rolled argument splitter instead of shlex.split. It is about reinventing code that exists elsewhere in the project or its dependencies, not about repetition inside one generated answer. On SAICA-Bench: No single best reviewer: 3 rows cannot be told apart at 95% on the samples they share (gpt-6-luna, mimo-v2.6-flash and deepseek-v4-flash); the highest J is +90.0% (95% interval +82.0% to +98.0%). Dependency blindness: results, incidents and tools →
Fabrication (code hallucination)
Code hallucination, which SAICA-Bench calls fabrication, is when an AI coding agent writes code that refers to a function, method, class or package that does not exist in any installed dependency, the standard library or the runtime. The code is syntactically plausible and factually false. On SAICA-Bench: No single best reviewer: 4 rows cannot be told apart at 95% on the samples they share (deepseek-v4.1-flash, glm-5.3, deepseek-v4-flash and gpt-6-luna); the highest J is +94.3% (95% interval +91.6% to +96.7%). Fabrication: results, incidents and tools →
Incomplete execution (false “done” claims)
Incomplete execution, also called premature or false completion, is when an AI coding agent reports a task as done while parts of it are skipped, stubbed out or silently dropped. Common forms are TODO comments left in the code, function bodies replaced with pass or ..., a feature only partly implemented, error paths skipped, or the agent stopping before every item the user asked for is addressed. On SAICA-Bench: No reviewer reaches J 25% on this set: the highest J of a row with enough readable answers is +19.5% (95% interval +13.9% to +24.9%), and 3 other rows cannot be told apart from it. No best reviewer is named. Incomplete execution: results, incidents and tools →
Logic error
A logic error is when an AI coding agent uses real, current APIs correctly but writes logic that computes the wrong result for the task, for example an inverted condition, an off-by-one or a faulty algorithm. It is the classic “passes the type check, fails correctness” bug. On SAICA-Bench: Best reviewer: mimo-v2.6-flash at J +74.3%. Its paired difference from the next row (deepseek-v4-flash) has a 95% interval that excludes 0. Logic error: results, incidents and tools →
Obsolescence (deprecated API usage)
Obsolescence is when an AI coding agent calls a real library API or package that has been deprecated, removed or materially changed, usually because the model’s training data predates the current release. The model saw an older version of the library and confidently writes the older form. On SAICA-Bench: No single best reviewer: 3 rows cannot be told apart at 95% on the samples they share (gpt-6-luna, mimo-v2.6-flash and deepseek-v4-flash); the highest J is +89.7% (95% interval +82.9% to +95.1%). Obsolescence: results, incidents and tools →
Scope creep
Scope creep is when an AI coding agent does more than it was asked: it edits files outside the request, installs packages nobody named, changes configuration outside the task, or adds refactors and “improvements” unprompted. It also covers committing unrelated changes alongside the requested edit and chaining follow-on work nobody asked for. On SAICA-Bench: Best reviewer: mimo-v2.6-flash at J +97.7%. Its paired difference from the next row (deepseek-v4-flash) has a 95% interval that excludes 0. Scope creep: results, incidents and tools →
Security vulnerability
A security vulnerability, as a coding-agent failure, is when an AI coding agent writes code that adds an exploitable flaw, such as an injection primitive, weak cryptography, a leaked secret or insecure deserialization, to code that otherwise looks correct. The flaw does not break the feature. On SAICA-Bench: No reviewer reaches J 25% on this set: the highest J is +7.9% (95% interval -0.0% to +15.9%), and 10 other rows cannot be told apart from it. No best reviewer is named. Security vulnerability: results, incidents and tools →
Supply-chain attack (slopsquatting)
A supply-chain attack through an AI coding agent is when the agent installs or invokes malicious code: a hallucinated package name that an attacker registered (slopsquatting), a near-miss of a popular name (typosquatting), or a compromised MCP server or tool. The agent can be led there by its own hallucination, by an adversarial package name or by a compromised upstream. On SAICA-Bench: No single best reviewer: 2 rows cannot be told apart at 95% on the samples they share (gpt-6-luna and deepseek-v4-flash); the highest J is +77.0% (95% interval +70.3% to +83.8%). Supply-chain attack: results, incidents and tools →
Test manipulation
Test manipulation is when an AI coding agent, faced with failing tests, changes the tests instead of the code under test, so the failures disappear while the bug stays. It is one form of reward hacking. Typical forms are rewriting assertions to match the current, wrong output, commenting out a failing test, adding pytest.mark.skip, replacing a real expectation with a tautology, or disabling the CI job. On SAICA-Bench: No single best reviewer: 8 rows cannot be told apart at 95% on the samples they share (minimax-m3, giskard, Qwen3.8-27B (self-hosted) and 5 more); the highest J is +91.0% (95% interval +81.3% to +94.9%). Test manipulation: results, incidents and tools →

Other names people use

Common names for the same failures, and which mode each belongs to.

Code hallucination
Code that refers to functions, methods, classes or packages that do not exist. API hallucination and phantom imports are the same failure; SAICA-Bench files all of them under fabrication. Fabrication →
Package hallucination
A model suggests importing or installing a package that does not exist. It is a form of fabrication; once an attacker registers the made-up name it becomes slopsquatting. Fabrication →
Slopsquatting
Registering a package name that AI models hallucinate, so that an agent which installs it runs the attacker’s code. One of the supply-chain attacks on coding agents; the best-known case is huggingface-cli. Supply-chain attack →
Typosquatting
Publishing a malicious package under a near-miss of a popular name (a doubled, dropped or swapped letter, a plural, a py- prefix) or a popular name plus a word, and waiting for a person or an agent to install it. A supply-chain attack; SAICA-Bench tests reviewers on known-malicious lookalike names. Supply-chain attack →
Reward hacking
Optimising the signal you are judged by instead of the goal behind it. In coding agents the usual form is making failing tests pass by changing the tests, which SAICA-Bench calls test manipulation. Test manipulation →
Premature completion (false completion)
An agent declaring a task finished while parts are skipped, stubbed or not working. SAICA-Bench’s term is incomplete execution. Incomplete execution →
Deprecated API usage
Code that calls a library API that has been deprecated or removed, usually because the model learned an older version of the library. SAICA-Bench’s term is obsolescence; training-data staleness is its usual cause. Obsolescence →
Reinvention (reinventing existing code)
Hand-writing a function the standard library, a dependency or the codebase already provides. SAICA-Bench’s term is dependency blindness. Dependency blindness →
Context rot
A common name for a model becoming less reliable as its context grows long. SAICA-Bench does not treat it as a separate mode: a session longer than the model can reliably attend over is one of the signals of context pollution. Context pollution →
Context poisoning
Deliberately planting false or malicious content in an agent’s memory or context; OWASP lists it as Memory & Context Poisoning (ASI06). Context pollution covers this and accidental drift as well. Context pollution →
Recovery spiral
A run of fix attempts in which each one introduces new errors. SAICA-Bench’s term is cascading failure; error cascade and fix regression are other names for it. Cascading failure →

Metrics and benchmark terms

How detection is scored and what the tool types mean.

J (J = TPR − FPR)
The score SAICA-Bench sorts by inside one failure mode. J = TPR - FPR, on the answers that carry a verdict. 0% is chance, 100% flags every bad sample and no clean one. J is the sort key within a mode; each row also has false alarms per 100 clean changes and precision at 1%, 5% and 20% prevalence. In statistics it is known as Youden’s J. J is not compared across modes: each set was built differently. How SAICA-Bench is scored →
TPR (true positive rate)
The share of bad samples a reviewer flags, also called recall or catch rate. A reviewer that flags everything has a TPR of 100% and is useless, which is why J subtracts the FPR.
FPR (false positive rate)
The share of clean samples a reviewer flags: its false-alarm rate. The tables also print it as false alarms per 100 clean changes. Where most changes are clean, the FPR decides how much reviewer time a check wastes.
Precision at a prevalence
Of the changes a reviewer flags, the share that are really bad, if a stated share of all changes were bad. It is computed from the TPR and FPR. The sets hold far more bad samples than real work does, so a reviewer with a high J can still raise more false alarms than true ones when failures are rare.
95% interval
95% intervals. TPR and FPR: Wilson. J: Newcombe where samples are independent; where samples are grouped (task, pair, API pair, file) whole groups are resampled (100,000 draws, fixed seed).
Answer with no verdict (unreadable)
An answer with no verdict (failed call, blocked request, text cut by the token budget, no verdict in the text) is unreadable: it is left out of the rates and counted beside them, per label. j_strict is J under the old rule that counted such answers as 'not flagged'; j_bounds is J if every unreadable answer had been wrong, and right.
Best reviewer, top group, best catalog tool
Best reviewer: the best ranked row of a mode that is not marked incomplete. A row with under 95% readable answers in one class is listed and ranked, with its J bounds, but is never named best reviewer or sole winner; if its J is higher than the best reviewer's it stays in the top group. Not ranked: baseline rules, shortcut baselines and rows outside a tool's design scope. Top group: the best reviewer and every ranked row whose paired difference from it includes 0. Best catalog tool: the best tool of the SAICA catalog that ran its own check; a framework that ran our rule text is not a candidate. Two rows are compared on the samples both have a readable verdict for; the difference of their J gets a 95% bootstrap interval (groups resampled where the mode has them). 'Ahead', 'worse' and a single winner need an interval that excludes 0 at the published precision: a bound that prints as 0.0% does not exclude 0. Leaderboard →
Mode-primed setting
Each model asked directly was told which failure to look for and was tested on samples that have that failure or none. Its score is a single-prompt, mode-primed estimate, not a product. Threats to validity →
Tiers
Tiers describe the best reviewer's J on our set, not how well a failure is caught in use: J 80% or more on our set; J 25% to under 80% on our set; J under 25% on our set.
Shortcut baseline
A rule that needs no understanding of the failure, for example one that reads only the task or the format of a patch. It is published as a row next to the reviewers to show how much of a score a set gives away for free. It is never ranked and never counted as best.
Cost per 1,000 reviews
A row's cost covers every call made to get its scored answers: the current answer of each sample plus earlier attempts that failed or were cut off and had to be re-run. Smoke tests, runs on earlier versions of a set and accidental duplicate answers are not included. Cost per 1,000 reviews is that cost divided by the reviews made, for one mode-specific question per review. 'Cheap' means at most $5.00 per 1,000 reviews. A “~” marks an estimated cost.
Sample (bad or clean)
One item to review: a task plus what an agent or a model produced for it, such as a patch, a function or a whole session. It is labelled bad if it contains the failure and clean if it does not; reviewers never see the label. Datasets and labels →
Paired samples (matched pairs)
A bad and a clean sample built from the same task, for example the same patch with and without a hand-rolled copy of shlex.split. Only some sets are built this way; each mode’s card on the benchmark page says how its samples were made.
Free program
A deterministic tool that needs no language model, such as a type checker, linter or security scanner. It costs nothing to run and gives the same answer every time. Run on SAICA-Bench: Ruff, Semgrep, Bandit, Pyright and mypy.
Packaged tool that wraps a model
A supervision tool that puts a language model inside its own review logic or judge template. Run on SAICA-Bench: Arize Phoenix, opik, DeepEval, pr-agent, giskard, promptfoo and uqlm, each with deepseek-v4-flash as its model, which is also tested on its own. Each row’s run card says whether the tool ran its own check or the rule text we wrote.
Model asked directly
A language model given the sample and one fixed prompt per mode that names the failure to look for. Each model asked directly was told which failure to look for and was tested on samples that have that failure or none. Its score is a single-prompt, mode-primed estimate, not a product.
Baseline rule
A few lines of logic written for the benchmark, for example flagging a patch that edits a test file. Baseline rules are shown for reference; they are not packaged tools and are never counted as best.
Supervision tool
Any tool that checks, constrains or records what a coding agent does: a reviewer, linter, scanner, sandbox, approval gate or tracer. Tools that detect a failure can be scored on SAICA-Bench; tools that prevent it or only record traces cannot be scored on a finished change. Tool catalog →